@muntasirfahim
AI & Backend Engineer
Track record
0
Projects done
New
Talent rating
Open
Availability
AI engineer who architects and owns production agentic systems end to end: orchestration, MCP tool platforms, RAG, Text-to-SQL, guardrails, evals, and deployment. Designed platforms serving 20 tenants and 2,000+ users , making the calls on security, safety, cost and latency and backing them with evidence. MSc (High Distinction), 10 peer-reviewed papers in explainable AI, former ML instructor.
AI Engineer L2
Surbana Technologies Pte. Ltd. · Remote (Singapore)
2025-12 – Present
• Architected an agentic analytics platform serving 20 tenants through chat, real-time voice, an embeddable SDK widget, dashboards and scheduled reports, with a LangGraph agent working across 6 custom MCP tool servers (SQL, semantic layer, RAG, charting). • Built a governed data layer: a Cube semantic layer as the default source, with data models generated from live schema introspection, and a Text-to-SQL fallback over PostgreSQL and MSSQL (temperature-0 single SQL author, code-enforced structural checks, an LLM reviewer for tenant rules and an automatic repair loop), so chat and voice return the same SQL for the same question. • Designed three-level multi-tenancy (tenant, sub-tenant, user) with resource inheritance, a per-resource sharing matrix and per-user bring-your-own-database connections, served by a connection pool manager sized for customer- owned databases. • Enforced zero-trust AI tool access through Agent gateway (RS256 JWTs, rotating JWKS, tenant-scoped claims) that injects caller identity as hidden tool arguments the LLM cannot see or spoof, with per-user row-level security in PostgreSQL, MSSQL and the Cube semantic layer that fails closed on pooled connections. • Cut agent initialization latency by 75% with parallel async loading and per-date agent caching; kept model context flat at any result size by passing query results between MCP servers by reference through a Redis artifact store and stripping bulk tool payloads from the model’s view. • Added human-in-the-loop clarification (LangGraph interrupts with resumable streaming) so the agent asks instead of guessing on ambiguous questions, plus a stream completion contract that detects cut-off turns, saves them to the thread, and tells them apart from a user’s Stop. • Designed a composable, database-driven guardrail pipeline (keyword filtering, LLM safety classification, Presidio PII redaction, conversation summarization) that tenants configure without code changes, extended to RAG ingestion. Stack: Python, FastAPI, LangGraph, LangSmith, AWS Bedrock, MCP, Cube, PostgreSQL, MSSQL, Redis, Next.js, Docker
Fahim Muntasir
Discuss your project
AI Engineer
Next Country · Remote (UK)
2025-03 – 2025-12
• Designed the architecture: a supervisor routes each case to 4 specialist visa agents (nomad, startup, tourist, employ- ment) over 4 PGVector knowledge bases, serving 2,000+ users . • Split the domain into specialist agents so each prompt stays small instead of one prompt covering every visa type. Stack: Python, LangGraph, FastAPI, Supabase, PGVector Agent Memory and Cost Control • Built long-term cross-session memory with automatic context pruning and optimized multi-turn tool execution for the NextBot chat app, cutting token consumption per run by 50% . • Implemented LLM Gateway that switches between pool of models, making the system fail-safe against high latency and model unavailability. • Introduced a relational per-tool usage ledger that attributes cost to every agent invocation. Stack: LangGraph, PostgreSQL, Supabase Visa Document Intelligence Pipeline • Shipped a pipeline covering crawling, multimodal LLM extraction, document generation and semantic indexing. • Added a human-in-the-loop review gate, cutting data collection time by 80% . Stack: Python, multimodal LLMs, PGVector, cloud storage
Machine Learning Engineer
LogicZen AI · Dhaka
2024-08 – 2025-03
• Built a pipeline over ball-by-ball match data with features for batting, bowling, pitch and match context. • Benchmarked XGBoost, LightGBM and neural networks as multi-class classifiers and tuned the best for deployment. Stack: Python, pandas, scikit-learn, XGBoost, LightGBM
AI & ML Instructor
Data People
2023-04 – 2023-11
designed and taught ML, deep learning and Python courses
ML Intern
Omdena
2023-01 – 2023-04
GIS-based fertilizer recommender with a Streamlit interface
Embedded Systems Intern
Metrosemi
2022-08 – 2022-12
anomaly analysis of chip-fabrication sensor data
BRAC University
MSc in Computer Science and Engineering
2023-09 – 2025-09
CGPA 3.78/4.0 (High Distinction). Thesis: a Bangla conversational AI agent for maternal health using MCP.
Rajshahi University of Engineering and Technology
BSc in Electrical and Electronic Engineering
2017-02 – 2023-05
Thesis: a soft-voting ML model with explainable AI for cardiovascular disease identification.
Japan Subsidy Bot
| Production client project 2025 Live 2 • Bilingual (Japanese/English) RAG assistant answering subsidy questions for call-center teams, live in production. • OAuth2 and RBAC accounts, per-IP rate limiting, response caching and LangSmith tracing on every request; React SPA and FastAPI API shipped as one Docker image. Stack: LangGraph, Claude Haiku 4.5, Gemini Embeddings, PGVector, FastAPI, React, Docker, LangSmith
View projectDischarge Letter Reconciliation
| Healthcare AI case study 2026 • Reconciles German hospital discharge letters against a nursing-home resident’s record; the LLM only extracts, while matching, conflict detection and routing to human review are deterministic code. • 0 hallucinated proposals in a labelled eval, 62 tests, chaos-tested failure modes, a risk register, about 1.5¢ per letter. Stack: Python, FastAPI, Pydantic, pytest
AI Credits: Usage-Based Metering and Billing
• Built usage-based billing that meters every model call across chat, voice, MCP tool servers and workflows into one credit ledger priced per model (including prompt-cache reads and writes), with daily, weekly and monthly caps per tenant and per user, Redis-backed counters, soft blocking, overage rates, statements and invoices. • Shipped an AI Credits admin portal with per-thread cost waterfalls (turn, tool call, model call) showing exactly where each credit went. Stack: Python, FastAPI, PostgreSQL, Redis, LangSmith, Next.js
Real-Time Multilingual Voice Agent
• Shipped a voice agent over LiveKit/WebRTC with two swappable engines (OpenAI Realtime speech-to-speech; Cartesia STT/TTS around Bedrock Claude) that shares the chat agent’s tools and tenant scoping and renders charts mid-call. • Kept spoken replies and on-screen chart and table labels in the caller’s language across 4 languages (English, Chinese, Malay, Tamil) by detecting the language in code (script detection plus a lightweight model check for Malay) and injecting it into tool calls instead of trusting the model. Stack: LiveKit, WebRTC, OpenAI Realtime API, Cartesia, AWS Bedrock (Claude), MCP
No-Code AI Workflow Studio
• Integrated the open-source ActivePieces engine natively into the Next.js portal as a no-code Workflow Studio (Node piece runtime, live run updates over Socket.IO, synchronous webhooks), with OAuth2 connections and PostgreSQL/MSSQL steps bound to each tenant’s connected databases. • Built its AI Agent step on MCP tools, integrations, and the tenant knowledge base, behind an OpenAI-compatible LLM proxy with bring-your-own-key providers, metered into the same credit ledger; left out arbitrary code-execution steps by design. Stack: ActivePieces, TypeScript, Node.js, Next.js, MCP, OpenAI-compatible LLM proxy
PR Guard
• Tool-calling review agent ( 16 git, GitHub CLI and filesystem tools) whose strict Pydantic output maps 1:1 onto the GitHub Reviews API, shipped as a CLI, a GitHub Action and a VS Code extension. • Runs on 5 model providers including local Ollama, so teams can keep code on their own machines; shipped with architecture docs, a QA guide and a rollout deck. Stack: LangGraph, LangChain, FastAPI, Typer, TypeScript, GitHub Actions, uv
