Melih – Python, AWS, AI agent development

• Company: A global IT engineering and consulting platform that provides digital transformation services, custom software development, and specialized staffing.
• Product: a code generation agent that translates unstructured requirements and technical documentation into executable simulation code using multi-agent orchestration.
• Users: Enterprise software developers, simulation engineers, and internal technical teams.
• Core Usage: Engineers upload functional specs or domain documents, and the tool automatically generates, tests, and refines executable simulation code, exposing outputs through a web interface.

• Partnered directly with client stakeholders to convert ambiguous, business-level requests into defined technical scope, running discovery sessions to establish what the system had to produce and under what constraints before committing to an architecture, then owning delivery through to production.
• Owned product scope rather than executing a fixed spec, deciding which agent capabilities shipped and which were cut.
• Ran discovery with the engineers who would actually use the tool, sequenced the roadmap into increments that could ship in a few weeks instead of a single large release, and dropped features that were not earning their maintenance cost once usage data came in.
• Architected and implemented agentic workflows for simulation-code generation, designing document-to-code pipelines that translated structured and unstructured inputs into executable outputs through multi-step LLM orchestration.
• Built and evaluated code-generation agents using custom agentic architectures, comparing performance, reliability, and developer ergonomics against Claude Code SDK to guide tooling and workflow decisions.
• Developed MCP tools and reusable Claude skills to extend agent capabilities, improve task specialization, and enable more structured human-in-the-loop and automated execution flows.
• Leveraged LangChain and custom orchestration patterns to coordinate planning, tool usage, context handling, and multi-stage reasoning across backend AI workflows.
• Built production-oriented backend services with FastAPI to expose agent functionality through clean APIs, supporting integration, testing, and scalable deployment patterns.
• Developed frontend interfaces in React to enable interaction with agent workflows, document ingestion pipelines, and generated code outputs.
• Deployed and operated AI services on GCP, aligning infrastructure with low-latency execution, service reliability, and practical experimentation needs across evolving agentic systems.

• Product: An end to end macroeconomic monitoring and policy support platform built for the Central Bank of Liberia, covering data ingestion from official CBL and international sources, FX and macro analytics, a multi model forecasting engine, automated early warning alerts, an LLM policy memo generator, and SURA, a tool calling research agent that sits on top of the whole system.
• Target Users: Central bank economists, monetary policy analysts, FX traders, and financial supervisors.
• Core Usage: Economists use SURA in natural language to execute multi-step statistical runs, trigger scenario shocks, generate policy memos with page-level citations, and track real-time early warning macroeconomic stress alerts through interactive React dashboards.
• Delivered as technical lead across AI/ML, backend, frontend, and infrastructure, and shipped complete against the full project plan.

• Acted as technical lead and primary technical counterpart to the central bank, running discovery with economists and policy staff to turn open ended supervisory and research needs into a concrete module roadmap, then owning architecture, scope decisions, and delivery of every module through to production.
• Built SURA, an agentic research assistant that plans and executes multi step economic analysis by routing across the platform's own tools for correlation, regression, forecast runs, scenario shocks, memo generation, and document search, with streaming responses, tool call traces, role gating, and refusal and citation guardrails. Benchmarked its planning and multi step execution quality against Claude Code and reached comparable reliability on chained analytical tasks, which made it the primary interface economists used instead of the individual dashboards.
• Designed and trained the forecasting engine across a wide model zoo, starting from ARIMA and SARIMAX, Prophet, and gradient boosted trees, then extending it with LSTM and Transformer baselines and later with Chronos time series foundation models and Bayesian VAR for structurally interpretable multivariate forecasting with proper uncertainty. Ran walk forward backtesting with blocked folds, calibrated 50, 80, and 95 percent prediction intervals, and registered the best model per target and horizon in MLflow.
• Shipped an LLM policy memo generator built on an agentic RAG pipeline over central bank documents, with layout aware PDF chunking, retrieval and re ranking, table extraction, automatic citation assembly with page level references and confidence scores, and guardrails that refuse non compliant statements and route low confidence output to human review. Held citation integrity above 95 percent under an evaluation harness, since a memo that misattributes a source is unusable in a policy setting.
• Developed the early warning system for macroeconomic stress, combining EWMA, CUSUM, and isolation forest detectors into a severity and confidence score backtested against historical episodes, tuned around the false positive and false negative tradeoff that decides whether analysts keep trusting the alerts, and wired to routed notifications with escalation rules and a full ingest to delivery audit trail.
• Built the FX and macro analytics hub end to end, exposing correlation and regression endpoints in FastAPI over a gold layer warehouse with input validation, confidence intervals, Redis caching, and sub two second p95 latency under load, paired with React views for the correlation matrix, regression diagnostics, and a virtualized data explorer for large series.
• Owned the data foundation the rest of the platform depended on, standing up a bronze to silver to gold pipeline for official FX rates, CPI, monetary aggregates, reserves and balance of payments, and fiscal data, with typed data contracts, layout aware PDF table extraction with a review queue, idempotent watermark based incremental sync, reconciliation against published figures, and Great Expectations suites plus freshness and null rate metrics enforcing data quality.
• Dealt with what live usage surfaced rather than stopping at launch, tracing bad or ungrounded model output back to retrieval and feature issues, tightening reconciliation where source data disagreed with published figures, cutting latency and inference cost on the paths analysts actually hit, and adding the prompt and tool tracing needed to answer why a specific forecast, alert, or memo came out the way it did.
• Made the scope calls on what shipped and what did not, cutting the design surface down to what the bank would use in its real reporting cycle and dropping work that was not earning its keep, which is why the full plan landed as a working system inside the timeline instead of a half finished build of everything.
• Deployed the platform on AWS with a versioned S3 and Postgres backbone, containerized services, GitHub Actions CI/CD with blue green releases and tested rollback, plus a fully parity tested on premise packaging so the bank could run the entire stack inside its own sovereign environment without external dependencies.

• Company: A Stockholm-based talent marketplace that connects businesses with pre-vetted, remote software developers and tech professionals.
• Product: The role involved architecting and deploying a multi-agent LLM architecture for an automated content generator.
• Target Users: Daily marketing teams, content strategists, and SEO managers within regulated industries.
• Core Usage: Marketers input campaign themes or target topics, and the platform scrapes real-time web data, enforces domain-specific regulatory rules, and automatically produces, enriches, and formats publish-ready digital content.

• Led technical discovery with client stakeholders across iGaming and betting platforms, translating open-ended business goals into concrete system requirements and prioritized scope, then designing, shipping, and iterating on the resulting solutions in production.
• Ran the content generation system in production with marketing teams using it daily, then owned what came back from real usage: traced bad or off-brand outputs back to retrieval and prompting failures and fixed them, cut latency and token cost on the hot paths as volume grew, and used structured tracing to answer stakeholders asking why a specific output came out the way it did.
• Architected and deployed a multi-agent LLM architecture for automated content generation, leveraging LangChain/LangGraph to orchestrate planner →agent → aggregator workflows. Incorporated advanced web scraping, search, and enrichment pipelines.
• Designed and scaled Retrieval-Augmented Generation pipelines combining Pinecone vector databases with Neo4j knowledge graphs to enable context-rich, real-time reasoning over complex data.
• Implemented custom inference pipelines tightly integrated with internal databases and deployed to AWS (EC2, Lambda, CI/CD) for low-latency, production-grade performance, empowering marketing teams with scalable post-generation workflows.
• Translated complex domain rules into machine-readable logic, creating structured representations for regulatory or policy-driven automation tasks. Built dynamic ETL pipelines for real-time synchronization between PostgreSQL and Neo4j.
• Fine-tuned and optimized generative models using PEFT and LoRA techniques to achieve domain-specific outputs with minimal resource overhead, suitable for regulated and high-precision use cases.
• Developed internal AI-powered enrichment tools for query handling, content summarization, document classification, and context-aware recommendation systems, accelerating knowledge discovery.
• Exposed robust LLM-based APIs for seamless integration into client-facing applications, emphasizing auth-first design, observability, and high availability.
• Added structured logging, tracing, and metrics for debugging and monitoring.
• Automated internal linking and relevance strategies, significantly improving SEO performance and content discoverability across digital platforms.
• Consistently operated with founding-tier ownership, translating ambiguous requirements into shipped features within days/weeks, aligning with startup execution velocity.

• Company: An AI platform focused on human intelligence and high-quality data used to train and improve frontier foundation models.
• Product: a production-grade AI-powered recruitment chatbot.
• Target Users: Job applicants and talent acquisition/recruitment teams.
• Core Usage: Candidates interact with the chatbot to ask complex questions about roles and compliance requirements, while recruiters rely on the system to parse applications, filter candidates, and run automated contextual screening.

• Worked directly with internal stakeholders and business owners to frame an unclear knowledge-access problem into a defined technical solution, establishing which queries the system had to answer and to what standard, then owning architecture and deployment end to end.
• Architected and developed a production-grade AI-powered recruitment chatbot leveraging Retrieval-Augmented Generation.
• Implemented a scalable agentic RAG pipeline incorporating Neo4j knowledge graphs for complex contextual reasoning and compliance-like query understanding.
• Designed advanced memory management systems, separating short-term conversation context from persistent long-term memory.
• Integrated sentiment analysis and adaptive response strategies to dynamically adjust the chatbot’s tone and escalation behavior.
• Engineered and optimized vector database architecture using Pinecone, employing Matryoshka Representation Learning to minimize storage footprint while maintaining high retrieval precision.
• Built and maintained dynamic ETL pipelines to synchronize data in real time between PostgreSQL and Neo4j, ensuring the knowledge graph remained continuously updated and production-ready.
• Led the design and deployment of robust, cloud-native infrastructure (AWS/GCP) for agent orchestration and retrieval, using Docker and CI/CD pipelines to achieve rapid, reliable releases.
• Created a real-time recommendation engine clustering users based on behavioral data and feeding results back into the pipeline.
• Collaborated with ML engineers and product stakeholders to translate domain-specific logic into machine-readable control flows.
• Instituted observability practices including structured logging, tracing, and performance metrics to monitor latency (p50/p95) and throughput, proactively improving system reliability.

• Product: A conversational intelligence platform designed for enterprises to analyze and improve human-to-human interactions, specifically within call centers and sales organizations. It uses AI to monitor 100% of customer communications—including audio, chat, and email—to drive better business outcomes.
• Target Users: Bank customer service representatives, call center agents, and enterprise sales teams.
• Core Usage: As a live call or chat takes place, the system transcribes the conversation in real time, automatically retrieves relevant customer data, and pops up dynamic desktop UI widgets to help agents solve customer inquiries faster.

• Created the architecture and implemented complex system pipelines utilizing Large Language Models (LLMs).
• Fintuned custom 10B+ LLM on bank specific dataset on multi-gpu system architecture.
• Implemented self-refining technique designed for the bank assistant tool.
• Enabled custom LLM to execute API calls, effectively retrieving necessary widgets and displays for the desktop application, thus enhancing system speed and performance.
• Integrated real-time processing and display capabilities in the desktop application, enhancing user experience through quicker response times.
• Used specialized knowledge in AI to design custom LLM-based pipelines, improving the quality and performance of our systems.
• Utilized innovative prompt engineering techniques to enhance the functionality and reliability of the AI systems.
• Integrated speech-to-text models for real-time transcription retrieval, streamlining communication and improving the overall performance of the bank assistant.
• Enabled LLMs to execute API calls, effectively retrieving necessary widgets and displays for the desktop application.
• Led the adoption of AI technologies in enhancing customer representative performance.
• Encouraged cross-functional collaboration among engineering, data science, and customer service teams.
• Led the development of a robust backend architecture to serve the desktop application.
• Provided regular updates to key stakeholders, including executive leadership and board members, on project progress, risks, and outcomes.
• Kept abreast with emerging technologies, market trends, and industry best practices.
• Led a team of engineers, fostering a collaborative and innovative work environment that promoted professional growth and knowledge sharing.

• Company: A data-driven technology company that uses an AI-powered platform to connect companies with remote software developers and tech talent from around the world.
• Product: A specialized conversational AI model fine-tuned on custom code datasets to understand, evaluate, and generate technical code snippets.
• Target Users: Global remote software developers and technical assessment teams.
• Core Usage: Used during candidate screening to evaluate technical coding queries, review applicant code submissions, and converse with developers on complex programming concepts in real time.

• Prepared code based dataset and train a GPT model to enhance the coding ability of the LLM side of the conversation model. Tried different finetuning strategies.
• Contributed to the development of the most popular LLM-based chatbot model having GPT architecture.
• Designed and implemented efficient and scalable algorithms for pre-processing, training, and evaluation of large-scale deep learning models.
• Developed and optimized software components to enable high-speed inference and seamless integration with various production systems.
• Developed novel techniques for transfer learning in deep learning models, specifically applied to the LLM-based chatbot model, resulting in improved performance and faster convergence.
• Implemented advanced natural language processing algorithms to enhance the chatbot model's understanding and generation of code-related queries and responses.
• Optimized memory utilization and runtime efficiency of the chatbot model.
• Researched and implemented state-of-the-art techniques, such as attention mechanisms and transformer architectures.

• Product: An automated medical image analysis platform that uses deep learning to detect and segment dental abnormalities from raw clinical X-rays.
• Target Users: Dentists, dental hygienists, and administrative claims review teams.
• Core Usage: Clinical staff upload patient X-rays, and the software automatically highlights cavities, cracks, and fractures, streamlining diagnostic reviews and accelerating administrative claims processing.

• Prepared and trained custom transformer networks for tooth detection, segmentation, keypoint detection, and various regression/classification problems using Pytorch.
• Trained models on Torch Elastic, Kubeflow, and AWS EC2 instances.
• Developed and implemented computer vision algorithms to analyze dental X-rays and detect abnormalities, such as cavities, cracks, and fractures.
• Designed and deployed a data pipeline to collect, preprocess, and augment dental X-rays from multiple sources, including dental clinics and hospitals.
• Optimized the performance of deep learning models by experimenting with different architectures, loss functions, and hyperparameters.
• Conducted rigorous testing and evaluation of models to ensure their accuracy, robustness, and generalizability on different datasets.
• Experimented extensively with a variety of custom neural network architectures and loss functions, leading to significant advancements in model performance and specificity for dental imaging applications.
• Conducted in-depth analysis and testing of various neural network configurations, identifying optimal structures for accurate detection of dental issues.
• Worked on making the AI models faster without losing accuracy.
• Converted Pytorch models to OpenVINO models for optimizing the models for the production environment.

Co-founded and served as CTO of a bootstrapped, fully self-funded startup that built synthetic data generation technology for banks and insurance companies, operating without any external investment.

• Product: An enterprise data platform that generates mathematically realistic synthetic datasets that mirror the exact statistical distributions of proprietary databases without containing any real personal information (PII).
• Target Users: Enterprise data science teams, software QA engineers, and compliance/risk officers in banking and insurance.
• Core Usage: Teams generate artificial, privacy-compliant datasets to build AI models, run software testing, and perform analytics without risking customer data privacy or waiting on lengthy manual PII scrubbing processes.

• Managed both the technical and commercial sides of the business, balancing hands-on engineering leadership with sales, delivery, and client success responsibilities.
• Led a team of 11 engineers and specialists, owning the full technical direction, architecture decisions, and product roadmap of the company.
• Developed a synthetic data platform that generated artificial datasets preserving the statistical properties, distributions, and relationships of clients' original data while containing zero real personal information.
• Eliminated privacy and compliance concerns for financial clients by replacing manual PII removal workflows with automated synthetic data generation, removing the risk of exposing real customer data.
• Replaced clients' existing manual PII deletion tools with a solution that produced ready-to-use datasets, cutting the time and effort teams spent on data anonymization.
• Enabled banks and insurance companies to safely use realistic data for development, testing, analytics, and model training without touching sensitive production records.
• Sold and deployed the product across multiple banks and insurance companies, building direct relationships with enterprise clients in a highly regulated sector.
• Drove the company to sustainable revenue purely through product sales, proving market demand without relying on venture funding or outside capital.
• Designed the synthetic data engine to maintain referential integrity and realistic data characteristics, so downstream systems and models behaved as if working with real data.