OPEN TO ENTRY-LEVEL OPPORTUNITIES

I build useful
AI applications.

I’m Jasir, a fresher AI/ML engineer in Kerala (B.Tech Computer Science and Engineering, 2022–2026). I build document QA workflows and ML inference APIs, and examine how well they actually work.

Kerala, IndiaAvailable immediatelyGitHub LinkedIn Hugging Face
01 / SELECTED SYSTEM JRUVA
USER QUESTION

Find the relevant context.

ADense retrievalBGE-small · Qdrant
BLexical retrievalOkapi BM25
RRFFuse → select context
Streamed answerGemini / Groq / OpenRouter
Architecture sketchSee the evaluation
MY WORKING STACKPythonFastAPIQdrantscikit-learnNext.js

01 / SELECTED WORK

The work.
The decisions behind it.

Three projects, from document retrieval
to model-serving contracts.

PUBLIC BETA2026

Jruva

A multi-provider assistant for document questions.

Hybrid retrieval, streamed responses and provider failover. Evaluated against dense and lexical baselines, with a close look at what the numbers mean.

FastAPINext.jsQdrantBM25 + RRF
Read the engineering notes

The system

FastAPI serves document-question-answering workflows to a Next.js/TypeScript interface. BGE-small embeddings in Qdrant and BM25 lexical retrieval are combined with reciprocal rank fusion. Retrieval is scoped to a user’s own documents.

The reliability work

Gemini, Groq and OpenRouter routing includes retries, circuit-breaker failover, response caching and per-user quotas. The project’s CI covers backend tests, migrations, linting and a TypeScript build.

A result that changed the story

On the recorded synthetic development benchmark, dense retrieval reached MRR 0.910; BM25 reached 0.810 and hybrid fusion 0.900. Adding another retrieval method did not improve the overall result.

What this does not establish

The corpus has only 50 synthetic passages. Ninety answerable queries were scored; ten unanswerable queries were excluded from these retrieval metrics. The same dataset was used during later tuning, so it is not an independent test of generalization or answer quality.

Timing, precisely

The median of about 5 ms wraps query embedding plus in-memory dense search on CPU, with the embedding model already loaded. It excludes model startup, HTTP/network time, reranking and LLM generation. Measured on an AMD Ryzen 9 7940HX; the run was not repeated. Earlier 0.36 ms hybrid timing used cached embeddings and is not comparable.

LEARNING PROJECT · REBUILDING2026

CustomerIQ

A churn-scoring API with explicit boundaries.

A Random Forest serving service with strict input validation, model-artifact checks and documented limitations. Built to expose what the model can—and cannot—claim.

scikit-learnPydanticGitHub Actions
Read the engineering notes

Serving contracts

Single-record and batch FastAPI endpoints serve scikit-learn preprocessing and Random Forest scores. Batch requests are bounded at 100 records. Pydantic rejects unknown fields, invalid categories, non-finite values and inconsistent service combinations.

Artifact integrity

Startup checks SHA-256 hashes, estimator metadata and feature order against a versioned manifest. A mismatch prevents the service from starting. Integrity verification does not establish predictive quality.

Configured verification

GitHub Actions runs pytest, Playwright browser scenarios, Docker builds, vulnerability scanning and authenticated inference smoke checks. These are configured checks, not a claim that every current run passes.

Model limitations

The legacy serving scores are uncalibrated. Training provenance is incomplete and the historical holdout was exposed in earlier work. Fresh labeled data is needed for an independent performance estimate; this demonstration is not approved for real retention decisions.

LEARNING PROJECT · REBUILDING2026

DomainTune

A QLoRA fine-tune for structured ticket triage.

Qwen2.5-1.5B fine-tuned to turn issue text into five JSON fields. On a 40-example gold set, valid schema-conformant answers rose from 22.5% to 92.5%, while error analysis showed what it has not yet learned.

PEFTTRL4-bit NF4LSH dedup
Read the engineering notes

Implemented scope

A Qwen2.5-1.5B QLoRA configuration uses PEFT, TRL, 4-bit NF4 and LoRA rank 16. The data code includes label mapping, locality-sensitive hashing for near-duplicate removal and 80/10/10 train/validation/test splits.

What the evaluation showed

Base and fine-tuned models were scored on the 40-example gold set (CPU, greedy decoding, first JSON object). The format was learned: valid outputs rose from 22.5% to 92.5%. The judgement was not: no output got all five fields right, sentiment is "neutral" for every ticket, priority is P2 for 90%, and components copy the project name.

Diagnosed failures and next steps

Only 1 of 40 outputs stops after the answer, which points to the end-of-sequence token missing from training targets. Next: add and unmask that token, rebalance labels, use a component taxonomy, then retrain and rerun. Training counts and logs were not recorded for the first run.

Evidence

Raw predictions, metrics and the evaluation scripts are published in the repository under eval_results/.

02 / A CLOSER LOOK

More components.
Not always better retrieval.

Explore a recorded development result
from Jruva’s retrieval benchmark.

RETRIEVAL COMPARISON
Dense0.910
BM250.810
Hybrid RRF0.900

Mean reciprocal rank · higher is better

50 synthetic passages100 total queries90 scored queries

DENSE / SEMANTIC SEARCH

0.910 MRR

The strongest overall result
on this development set.

BGE-small embeddings and Qdrant retrieved the relevant passage within the top three for 98.9% of the 90 scored queries.

The median of about 5 ms covers CPU query embedding and in-memory dense search only. It is not end-to-end answer latency.

Scope matters. One relevant passage per answerable query. Ten unanswerable queries are excluded from these scores. Synthetic corpus, shared dataset/code authorship and tuning on the same queries; no independent holdout. These figures describe development retrieval behavior, not production answer accuracy. Rerun the benchmark: dataset and script

03 / THE PERSON BEHIND THE WORK

Curious about the model.
Careful with the claim.

I studied B.Tech Computer Science and Engineering at MES College of Engineering, Kuttippuram (2022–2026). My focus is the application layer: making retrieval, model serving and interfaces work together.

I use AI coding agents substantially in my projects. The engineering notes above set out what each project does, and what its evidence does not yet establish.

Connect on LinkedIn
01

LLM applications & retrieval

Document QA, embeddings, vector search, BM25, RRF and provider routing.

02

ML inference & contracts

scikit-learn serving, FastAPI, Pydantic, manifests and honest model metadata.

03

Interfaces & delivery

Next.js, TypeScript, GitHub Actions and automated checks.

EDUCATION & TEAM EXPERIENCE

B.Tech · Computer Science & EngineeringMES College of Engineering · 2022–2026 · Course completed, one result awaitedNASA Space Apps Challenge 2024 · Global Nominee and People’s Choice

Website developer in six-member Team DarkNova (Thrissur local event) for the Galactic Games challenge.

Official team page Team project source

AVAILABLE IMMEDIATELY

Have an AI or Python
role in mind?

I’m looking for an entry-level opportunity to build, evaluate and improve AI applications with a team.