AI & LLMs
- LangChain
- LangGraph
- Deep Agents
- Ollama
- FastMCP
- MCP
- Agent Workflows
- Tool Calling
- Prompt Engineering
Sajid MiyaAI Engineer
BUILDING WHAT COMES NEXT
I build agenticsystems thatsurvive contactwith production.
Three full-stack AI products — SaviraNous, NexHire, and commit2connect-mcp — plus Portfolio Agent, a local-first assistant. HiPPA evaluates pronunciation at three levels. Stack: Python, FastAPI, LangChain, LangGraph, and FastMCP.
Open to AI / LLM engineer roles · India, Nepal, remote.
0
AI projects built, end-to-end
0.0 yr
Building agentic systems in production
0.000
Sentence fluency Pearson (HiPPA)
About
The short version.
I build agentic systems, RAG pipelines, and LLM-powered applications. My portfolio includes three full-stack AI products — SaviraNous, NexHire, and commit2connect-mcp — plus Portfolio Agent, a separate local-first assistant.
My work spans multi-agent orchestration, MCP integrations, and systematic evaluation with DeepEval and Ragas. I build with Python, FastAPI, LangChain, LangGraph, and vector databases including FAISS, ChromaDB, and Pinecone.
Alongside application engineering, I work with PEFT, LoRA, and QLoRA fine-tuning. My HiPPA research uses a reference-conditioned WavLM transformer to assess pronunciation at phoneme, word, and sentence levels. I start with a working system, refine it relentlessly, and focus on production readiness.
“Start dirty, refine relentlessly. Ship what works. Polish what matters. Make it production-ready.”
Capabilities
What I do.
The verbs come first. Tools are how I get there.
RAG & Evaluation
- RAG Pipelines
- FAISS
- ChromaDB
- Pinecone
- Cross-Encoder Reranking
- DeepEval
- Ragas
Fine-Tuning & ML
- PEFT
- LoRA
- QLoRA
- Continued Pretraining
- Supervised Fine-Tuning
- Instruction Tuning
- DPO / RLHF
- Scikit-learn
- Pandas
- NumPy
- TensorFlow
- PyTorch
- NLP
Languages & Backend
- Python
- C++
- C
- JavaScript
- SQL
- FastAPI
- Flask
- Flask-SocketIO
- REST APIs
Data & Cloud
- PostgreSQL
- MySQL
- MongoDB
- Redis
- SQLite
- Docker
- AWS
- Git
- GitHub
Stack
Python · FastAPI · LangGraph
Fine-tuning
PEFT · LoRA · QLoRA
Data
PostgreSQL · MySQL · Redis · SQLite
Delivery
Docker · AWS · GitHub
Field Record
What I've shipped.
Four solo-built AI projects. Reverse chronological.
SaviraNous
Creator · Conversational AI Tutor · 2026 · Independent
Desktop conversational AI tutor with secure multi-user workflows, PDF and web RAG, four-layer memory, and persistent conversations.
- DeepEval across 38 curated cases
- 0.990 answer relevancy · 0.978 faithfulness
- 0.925 contextual recall
- Four-layer conversational memory
- Built secure authentication and conversation management for multiple concurrent users.
- Built PDF and web RAG with FAISS and ChromaDB, resource-aware retrieval, and cross-encoder reranking.
- Evaluated 38 curated cases with DeepEval: 0.990 answer relevancy, 0.978 faithfulness, and 0.925 contextual recall.
- Implemented four-layer memory, checkpoint restoration, conversation branching and merging, persistent threads, and an admin dashboard.
Python, FastAPI, PostgreSQL, Electron, FAISS, ChromaDB, DeepEvalOpen live
NexHire
Creator · AI-Native Hiring Platform · 2026 · Independent
End-to-end recruitment platform — job posting through offer — with LLM-driven screening, matching, scheduling, and email.
- 4 third-party API integrations
- Subscription billing live
- Multi-tenant data isolation
- Automated resume screening, candidate matching, interview scheduling, and email generation using LLMs, RAG, and agentic workflows.
- Integrated Google Calendar, Gmail, Slack, and Stripe for scheduling, communication, team notifications, and subscription billing.
- Stack: React, FastAPI, PostgreSQL, Redis, Docker, JWT.
commit2connect-mcp
Creator · MCP Server Bridging GitHub + LinkedIn · 2026 · Independent
A single FastMCP server connecting GitHub repository management and LinkedIn publishing through one MCP connection, with 56 tools, 26 prompts, and 4 resources.
- 56 tools · 26 prompts · 4 resources
- Long-running task support
- 30s cache + rate-limit guard
- Built a FastMCP server connecting GitHub repository management and LinkedIn publishing through one MCP connection, with 56 tools, 26 prompts, and 4 resources.
- Added tool annotations, response caching, rate limiting, and long-running media uploads.
- Connected it to Claude Code and Codex, including a workflow from code changes and pull-request merge to LinkedIn announcement.
Python, FastMCP, uv, OAuth 2.0, GitHub REST API, LinkedIn API
Portfolio Agent
Creator · Local-First AI Assistant · 2025 · Independent
Privacy-first AI assistant. Zero external API dependency. Fully local multi-modal with human-in-the-loop approvals.
- Fully local — zero cloud calls
- Human-in-the-loop approvals
- Vision + tool calling
- Persistent long-term memory
- Built with FastAPI, LangGraph, Ollama, and SQLite — real-time streaming chat with no cloud calls.
- Designed persistent SQLite + JSON long-term memory for coherent context across sessions.
- Added vision input, tool calling, dynamic model orchestration, and explicit HITL checkpoints for sensitive actions — fully local multi-modal deployment.
FastAPI, LangGraph, Ollama, SQLite
Research
HiPPA.
Hierarchical multi-task transformer for pronunciation assessment on SpeechOcean762 — extending the benchmark's coverage beyond its phoneme-only baseline.
- Venue
- SpeechOcean762 benchmark
- Task
- Pronunciation Assessment
- Architecture
- Hierarchical MT-Transformer
- Framing
- Beyond phoneme baseline
HiPPA — Hierarchical Multi-Granularity Pronunciation Assessment Pipeline
Built a reference-conditioned WavLM transformer for SpeechOcean762 that scores pronunciation at phoneme, word, and sentence levels, with multi-task learning, cross-attention, CTC-based confidence features, and score calibration.
The full pipeline covers preprocessing, training, evaluation, report generation, and a Streamlit analysis dashboard.
Read the paper (PDF)0.000
Sentence fluency Pearson
+67% relative to the baseline (0.450 → 0.754)
Sentence prosody Pearson0.737
Sentence total Pearson0.697
The trade-off: a small drop on phoneme-level scoring (0.450 → 0.396), in exchange for three new metrics the benchmark did not previously support.
Credentials
Schools & scholarships.
The academic record behind the work.
2023 — 2027
Thapar Institute of Engineering and Technology
B.E. Computer Science and Engineering
Patiala, Punjab · 8.68 / 10 CGPA
2021 — 2023
St. Xavier's College, Maitighar
School Leaving Certificate (SLC)
Kathmandu, Nepal · 3.66 / 4 GPA
2021
The Old Capital Secondary School
Secondary Education Examination (SEE)
Raniban, Gorkha, Nepal · 3.95 / 4 GPA
COMPEX Scholarship — Government of India (EdCIL)
2023 — Present
Competitively awarded merit scholarship for Nepalese students studying in India. Covers full tuition and hostel at Thapar.