$ whoami

Samarth Pratap Singh — B.Tech CSE, VIT Bhopal ('23–'27) · CGPA 8.57
AI/ML Engineer Intern @ AmberFlux EdgeAI

I build the boring-but-hard parts of AI products: agent orchestration
that doesn't fall over, retrieval that's actually measured, and memory
that persists across sessions. Every project below ships with numbers,
not adjectives.

$ neofetch

        ▗▄▄▄▖               samarth@vit-bhopal
       ▐▓▓▓▓▓▌              ─────────────────────────────
      ▐▓▓▓▓▓▓▓▌             OS ............ VIT Bhopal, CSE '27
   ▄▄▄▝▜▓▓▓▓▓▛▘▄▄▄          Host ........... AmberFlux EdgeAI (intern)
  █▓▓▓▓▄ ▝▀▀▀▘ ▄▓▓▓▓█       Kernel ......... LangGraph + FastAPI
  █▓▓▓▓▓▓▄▄▄▄▄▓▓▓▓▓▓█       Shell .......... python3 --strict
   ▀▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▀        Uptime ......... 240+ DSA · 4 shipped pipelines
     ▀▀▜▓▓▓▓▓▓▓▛▀▀          CPU ............ caffeine (98% util)
        ▝▀▀▀▀▀▘             Memory ......... Postgres · Qdrant · MongoDB
                            Theme .......... Signal Block [neubrutalist]
                            Status ......... seeking full-time, Aug 2026

$ cat capabilities.json

{
  "orchestration": ["LangGraph", "multi-agent supervisors", "conditional routing"],
  "retrieval": ["hybrid search (BM25 + dense)", "RRF fusion", "cross-encoder rerank"],
  "memory_protocol": ["MCP / FastMCP", "Postgres checkpointing", "Qdrant semantic recall"],
  "serving": ["FastAPI", "Next.js", "Docker"],
  "eval": ["LLM-as-Judge", "Ragas", "ROUGE / BERTScore / METEOR"],
  "core_ml": ["PyTorch", "Transformers", "PEFT / LoRA"],
  "observability": ["LangSmith", "MLflow", "W&B"],
  "languages": ["Python", "C++", "TypeScript", "SQL"]
}

$ ls pipelines/ --status

01. doc_copilot/ RAG document Q&A — hybrid retrieval (BM25 + dense) with RRF fusion and cross-encoder rerank.

correctness   89.2%   (+31 pts over keyword-matching baseline)
guardrails    prompt-injection detection · PII redaction
stack         Qdrant · Groq Llama-4-Scout · Next.js · FastAPI

→ github.com/noviciusss/DoCopilot

---

02. argus/ Multi-agent research engine — supervisor routes planner → researcher → critic → writer.

turnaround    30–90s   (was: hours, done manually)
control       critic agent auto-rejects weak drafts and re-routes
tracing       fully traced in LangSmith, async submit → poll → fetch

→ github.com/noviciusss/Argus

---

03. contextcore/ Stateful memory agent with a custom FastMCP server exposing tools over MCP.

memory        Postgres checkpointing + Qdrant recall + MongoDB profiles
router        intent-based, conditional LangGraph edges — no mode-switch
tests         17/17 passing

→ github.com/noviciusss/ContextCore-CLI

---

04. dialogue_summarizer/ LoRA fine-tune of FLAN-T5-base on SAMSum — 2% of parameters updated.

rouge-1       49.01     bertscore_f1   72.25     meteor   42.51
result        matches full fine-tuning at a fraction of the compute

→ huggingface.co/spaces/noviciusss/dialogue-summarizer

---

05. agent_guard/ AST-based static analysis for agent code — catches unbounded retry loops, unsafe shared state in fan-out branches, and timeout-less LLM/HTTP calls. CLI + GitHub Action. No ship date, no pressure.

$ cat model_card.yaml

model_name: samarth-pratap-singh
version: v4.2027-final-year
architecture: human · caffeinated · stubborn about eval numbers
parameters: underestimated (probably)
training_data:
  - 240+ leetcode problems (DP · graphs · trees — still hardening)
  - 4 shipped agent/RAG pipelines with real benchmarks
  - 1 internship, 2 codebases, 0 patience for vague specs
known_limitations:
  - will not ship a metric it hasn't personally verified
  - allergic to "production-grade" claims without a database to back it
intended_use: SDE / AI-ML full-time roles, 2027 batch
license: open-to-work

$ curl /telemetry

$ connect --to=samarth

$ off_duty --list
gaming · edm / lo-fi / metal · manhwa · sci-fi & thrillers

_{process exited 0 · thanks for reading the logs}