_{01 ABOUT  ·  02 IMPACT  ·  03 TECH STACK  ·  04 EXPERIENCE  ·  05 FEATURED PROJECT  ·  06 EDUCATION}

🧠 GenAI & ML Stack
☁️ Cloud & Orchestration
💻 Languages & Frameworks
📚 Data & ML Libraries
🧪 Methodologies
📊 Data Visualization

🔬 Featured Project — Gemma 3 1B Reasoning Fine-Tuning with Tunix

LoRA SFT → GRPO

Strict XML Output Contract • Kaggle TPU–stable • Judge-ready inference wrapper

This project documents a production-grade reasoning fine-tuning pipeline for Gemma 3 1B Instruct using Tunix (JAX) on Kaggle TPUs. The core goal is judge robustness: ensure the model emits a machine-parseable output schema every time:

<reasoning>...</reasoning>
<answer>...</answer>

• Stage 1 — LoRA SFT: Seeds strict XML format + instruction-following behavior.
• Stage 2 — GRPO: Reinforcement learning with composite rewards (format + task quality + stability).
• Judge-safe inference wrapper: Deterministic repair + escaping guarantees strict XML for evaluation.
• Systems reliability: OOM-safe GRPO defaults, KV-cache sizing guarantees, and Tunix 0.1.6 concat-safe rollout batching.
• Evaluation harness: Quant metrics, qualitative before/after, and TensorBoard-derived training curves.

Languages:   Python

These are small-N smoke tests designed to catch regressions early (format breaks, empty outputs, instability). The key win is eliminating judge-visible format failures via deterministic strict-XML repair.

FieldScribe is an offline‑first, evidence‑anchored triage copilot that: (1) localizes chest X‑ray findings, (2) detects interval change (Δ heatmap), (3) compresses FHIR‑shaped EHR context into auditable facts, and (4) exports a clinician‑readable triage report + tamper‑evident PDF.

• Offline-first: Runs on-prem / edge once weights & assets are present (no cloud dependency at inference time).
• Evidence-first UX: Always renders images + Δ heatmap + extracted facts alongside the narrative output.
• MedGemma-powered: Uses MedGemma 1.5 4B IT for multimodal synthesis (CXR + EHR).
• Safety guardrails: DICOM routing interlock prevents patient mismatches from entering synthesis.
• Auditability: Structured JSON audit packet + Unicode-safe PDF export (ReportLab).
• Feasibility instrumentation: Logs latency, peak VRAM, and token counts (ablation table in notebook).

• Time saved per complex triage case: ~10–14 minutes (manual EHR review + prior comparison + note drafting → evidence-first workflow).
• Capacity reclaimed (example): 40 cases/day × 10 min/case ≈ 6.7 clinician hours/day returned to care.
• Access + privacy: Enables multimodal assistance in bandwidth-limited or sovereignty-restricted settings.

In many clinical environments, the bottleneck isn’t “can AI answer medical questions?” — it’s last‑mile triage under severe operational constraints:

• Intermittent connectivity: rural clinics, mobile units, disaster response, overcrowded EDs can’t rely on cloud uptime.
• Strict data sovereignty: PHI often cannot leave the site or cross borders.
• Fragmented evidence: images, EHR facts, and priors exist — but clinicians must fuse them under intense time pressure.

FieldScribe targets a concrete workflow: longitudinal chest X‑ray (CXR) interval change triage where a clinician must:

1. detect what changed,
2. localize it,
3. reconcile with EHR context, and
4. write an auditable note.

Status: Research / demo prototype.
Safety: Not a medical device. Outputs require clinician verification.

This README expects the following files under assets/:

• Multimodal_Triage.jpeg
• Clinical_Pipeline_Architecture.png
• Multimodal_Triage_Dashboard.png
• Agent_Demo.jpeg
• Longitudinal_CXR_Delta_Heatmap.png
• Safety_Guardrails_and_Audit_Trace.jpeg
• Auditable_Clinical_Summary_Report.pdf

➡️ Download the example PDF: Auditable_Clinical_Summary_Report.pdf

• Accepts an offline FHIR‑shaped EHR bundle (JSON).
• Runs a Fact Filter stage that compresses dense EHR into query‑relevant bullet facts (non‑inventive, audit‑friendly).
• Goal: keep the LLM context window clean, reduce hallucinations, preserve traceability.

• Computes an interval change heatmap (current vs prior CXR).
• Produces a visual evidence panel clinicians can inspect immediately (value even before text generation).

• Uses MedGemma 1.5 4B IT as the multimodal model for:
  • longitudinal imaging review
  • anatomical localization / grounding
  • EHR‑informed synthesis
• Runs deterministic inference (do_sample=False) for reproducibility; optionally exposes a “glass‑box” trace in the UI.

• Generates an executive triage summary and a FHIR‑shaped audit packet.
• Exports a Unicode‑safe PDF (ReportLab) with evidence images, trace metadata, and disclaimers.

• Offline‑first design: works without external service calls once weights/assets are present.
• Run‑all notebook reproducibility: deterministic settings + auto‑reset YAML config pattern.
• Safety interlock: DICOM header routing prevents mismatched patient context from entering synthesis.
• Performance instrumentation: captures latency, output tokens, and peak VRAM across ablations.
• Clinician-style UX: evidence surfaced alongside narrative output.

• Target: single Tesla T4‑class GPU (edge deployable).
• Includes:
  • Warmup after model load to reduce first‑click latency spikes.
  • Fast Mode to cap tokens / reduce context while preserving evidence‑first outputs.
  • Ablation table logging: latency (s), output tokens, peak VRAM (MB)

Example observed in demo screenshot: Peak VRAM during visualization ≈ 3092 MB (heatmap + panel rendering).
Exact measurements depend on runtime + model caching; see the notebook’s feasibility section.

FieldScribe is a demonstration prototype. It is not a regulated medical device and must not be used for definitive diagnosis or treatment decisions. Always defer to qualified clinical judgment and local protocols.

Challenge	Models	Ranking	People Ranking
House Price Prediction	Stacking Models	Top 3%	Top 97 / 4651
House Price Prediction	XGBoost, LightGBM	Top 12%

▪️ Predict whether a customer will churn using machine learning models

▪️ Covid 19 — Data Analyst Mexico Overview

▪️ Explanation model of death cases in Mexico

▪️ Mexico Counties Covid Cases Correlation

▪️ Mexico`s Mobility pattern behaviour

▪️ Mexico Counties Clustering

▪️ XGBoostRegressor

▪️ RandomForestRegressor

▪️ DecisionTreeRegressor