AI engineer building agentic systems, document intelligence, semantic search, and the occasional programming language—because apparently one syntax wasn’t enough.
I build AI systems that have to work outside a notebook: RAG pipelines, ReAct agents, search infrastructure, document parsers, and FastAPI services that survive contact with real users.
Some people collect browser tabs. I collect retrieval metrics.
| Signal | Result |
|---|---|
| RAG answer accuracy | 93% |
| Domain embedding recall | 96% |
| Active-user growth linked to NLP personalization | 2× |
| Longer average sessions | 120% |
Parse documents in the browser, at roughly 100 pages per second, without sending them anywhere.
A privacy-first document intelligence engine built with TypeScript, LightGBM, ONNX Runtime Web, WebAssembly, and OpenCV. It runs fully on-device, understands document layout, and turns PDFs into structured, RAG-ready content—even offline or air-gapped.
91.34% accuracy · 86.32% macro F1 · CPU only · No external API
An AI copilot for electrical networks ranging from 14 to 70,000+ buses.
The platform combines React, FastAPI, LangChain, PyTorch, MongoDB, a graph neural network, and a ReAct agent armed with 60+ tools for topology, power flow, contingency, OPF, and SCADA queries.
60+ agent tools · 100+ MATPOWER cases · RAGAS scores: 1.00
A Unicode-native programming language written in Devanagari.
Custom lexer, recursive-descent parser, type-safe AST, tree-walking interpreter, closures, objects, arrays, localized errors, REPL, and a VS Code extension. It started as a compiler project and escalated responsibly.
TypeScript · Compiler design · देवनागरी · Yes, it actually runs
AI & NLP LLMs · RAG · ReAct · embeddings · reranking · RAGAS
ML PyTorch · Transformers · SentenceTransformers · LightGBM · YOLO
Search Apache Solr · BM25 · dense search · hybrid search · RankLib
Backend Python · FastAPI · MongoDB · MinIO · Keycloak
Frontend TypeScript · JavaScript · React · ONNX Runtime Web · WebAssembly
Infrastructure Docker · Docker Compose · Azure · Git
- Give the model tools, context, and a way to admit uncertainty.
- Measure retrieval before blaming generation.
- Keep the clever part explainable and the boring part reliable.
- Ship the demo. Benchmarks are nicer when they have a URL.
Building production AI systems at 1Ansah Technologies, with a particular interest in retrieval, agent orchestration, document AI, and search that understands what the user meant—not just what they typed.