Debashish Thakur
AI/LLM Engineer
Overview
Social links
About
AI/LLM engineer with 3 years building and shipping production-grade agent systems.
- Until October 2026 I led agent engineering at BlueBear AI, building multi-agent workflows, MCP servers and structured-output pipelines for B2B automation.
- I write the guardrails too — skill- and rule-based access controls that decide what an agent may touch before it touches it.
- As founding ML engineer at Ayna AI I ran an LLM/VLM analytics framework across 10K+ workflow outputs and cut error rates by 18%.
- I rebuilt their image pipeline to serve 3K+ requests a day at 40% faster inference and 30% lower infrastructure cost.
- Most of my time goes to the unglamorous half of the problem: output that holds its shape and tool calls that fail loudly instead of silently.
Stack(21)
TypeScript
C++
LangChain
Claude
OpenAI
MCP
Hugging Face
scikit-learn
ComfyUI
FastAPI
Django
Flask
Docker
PostgreSQL
MySQL
AWS
Google Cloud
n8n
Jupyter
Git
Experience(3)
Founding AI/LLM Engineer
- Led development of production-grade LLM agents for B2B use cases, building MCP servers and structured-output pipelines.
- Orchestrated multi-agent workflows for complex automation, coordinating task delegation and execution across agents.
- Implemented security guardrails — skill- and rule-based access controls that constrain agent behavior and block unauthorized actions.
- Applied systematic prompt engineering to raise agent output accuracy and consistency.
- Python
- TypeScript
- MCP
- Anthropic API
- Docker
- Postgres
- nsjail
Founding ML Engineer
- Architected and deployed an LLM/VLM analytics framework processing 10K+ workflow outputs across text and images, surfacing 70+ edge cases and cutting error rates by 18%.
- Built a ComfyUI-based image generation pipeline serving 3K+ daily requests — 40% faster inference and 30% lower infrastructure cost.
- Led MCP server adoption for enterprise LLMs, automating three internal workflows and speeding up cross-team knowledge retrieval.
- Engineered the dataset pipeline for a multiclass segmentation model trained on 50K+ images, refining the set with vision-model inference.
- Python
- PyTorch
- ComfyUI
- MCP
- Vision models
- AWS
Inspiron Info Inc
Machine Learning Engineer
- Designed and validated a movie revenue forecasting model in scikit-learn at 87% accuracy — 12 points above baseline.
- Engineered an LLM-powered text-to-SQL engine enabling natural language database queries with Matplotlib, Plotly and Seaborn visualization.
- Trained LSTM-based sequence models from scratch for natural language to SQL translation.
- Python
- scikit-learn
- LSTM
- SQL
- Plotly
Projects(5)
Second Brain
Portable memory for AI agents: an Obsidian vault served over MCP, on npm and the official MCP Registry, rated A on Glama.
- secor shiplocal1234Highsecret in git history567Medany-origin CORS89101125rulescore5api9mcp5web6
SECOR
Local security checks for what you build with AI, and a guard on every MCP tool call your agent makes.
M8ven 89/1001,368 tests10.2026 — Present - mail.google.comलेRewriteShortenTranslateAskClaudeGPTGeminiNanoyour keyCopyReplace
LekhakAI
A Chrome extension that sends one prompt to several LLMs from anywhere in the browser.
01.2025 — Present
- a prompt that turns plain English into SQLby meaning≈6,000 prompts, embedded6 nearest by meaning
The Prompt Craft
Semantic search over roughly 6,000 prompts gathered from Reddit.
05.2025 — Present - brief.pdfPDFterms.pdfPDFscan.pdfPDFmerged.pdfPDFMergeSplitCompressConvertnothing uploaded
meanpdf
A local-first PDF toolkit: merge, split, compress and convert without uploading a page.
02.2026 — Present
Open source(11)
Openfix(beliefs): read contradicts edges so belief derivation sees contradictions
doobidoo/mcp-memory-service · #1692 · 26.10.2026
- Issue open
Belief derivation never sees contradictions: _find_contradictions() reads storage.graph, calls a missing get_edges() and asks for an edge type nothing writes
doobidoo/mcp-memory-service · #1657 · 23.10.2026
- Issue fixed
Decay connection boost never applies on SQLite-Vec and Hybrid (get_memory_connections returns tag keys, decay looks up content hashes)
doobidoo/mcp-memory-service · #1592 · 19.10.2026
- Merged
fix(consolidation): key SQLite memory connections by content hash so decay boosts linked memories
doobidoo/mcp-memory-service · #1595 · 19.10.2026
- Issue fixed
Hybrid backend: mode="hybrid" silently runs vector-only search
doobidoo/mcp-memory-service · #1576 · 17.10.2026
Show 6 moreShow less
- Merged
fix(hybrid): run mode="hybrid" on the hybrid backend
doobidoo/mcp-memory-service · #1581 · 17.10.2026
- Issue fixed
Hybrid search: BM25 normalization gives the best keyword matches the lowest score
doobidoo/mcp-memory-service · #1567 · 17.10.2026
- Merged
fix(search): give stronger BM25 matches the higher keyword score
doobidoo/mcp-memory-service · #1568 · 17.10.2026
OpenAdd opt-in human feedback for selection and reflection
google-research/harness2 · #6 · 16.10.2026
- Issue open
Proposal: opt-in human feedback for selection and lessons (off by default)
google-research/harness2 · #2 · 16.10.2026
OpenAdd debashishthakur/brain-mcp to Knowledge & Memory 🤖🤖🤖
punkpeye/awesome-mcp-servers · #15994 · 08.10.2026
Education
Assam Engineering College
B.Tech, Computer Science and Engineering · Guwahati, Assam
Awards(3)
Smart India Hackathon Finalist
2022Top 5 nationally. Led a team of five to build a real-time analysis tool for hackathon judges.
Inter-college robotics champion
2022Architected an autonomous obstacle-avoiding robot that beat a field of 20+ teams.
Coding Club Secretary
2021Ran 10+ technical events for 200+ students and grew sustained club participation.