Skip to content
View noviciusss's full-sized avatar
💭
studying
💭
studying
10 followers · 19 following

Achievements

Achievement: Pull Shark

Achievements

Achievement: Pull Shark

Highlights

  • Pro

Block or report noviciusss

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
noviciusss/README.md
terminal typing

static static static


$ 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

Email Portfolio LinkedIn HuggingFace


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

process exited 0 · thanks for reading the logs

Pinned Loading

  1. DoCopilot DoCopilot Public

    RAG-powered document Q&A with 89% accuracy. Upload PDFs, ask questions, get cited answers. Built with LangChain + Qdrant hybrid search (BM25 + Vector) + Cross-Encoder reranking + Groq LLM. Includes…

    Python 2

  2. Argus Argus Public

    A production-grade multi-agent research pipeline that autonomously plans, researches, critiques, and synthesizes comprehensive cited reports from any research query -- with real-time streaming logs…

    Python 1

  3. portfolio_ts portfolio_ts Public

    My Official Portfolio :)

    TypeScript 1

  4. LeetCode_solutions LeetCode_solutions Public

    A collection of LeetCode questions to ace the coding interview! - Created using [LeetHub v2](https://github.com/arunbhardwaj/LeetHub-2.0)

    C++ 1

  5. Deep_learning Deep_learning Public

    Jupyter Notebook 1

  6. FineTunning FineTunning Public

    Jupyter Notebook