SECTION: 004 // LAB & R&D PROTOTYPES·DIR: /EXPERIMENTS/PROTOTYPES

COORDS: 33.6844° N, 73.0479° E·STATUS: UNCONSTRAINED

APPLIED R&D & TECHNICAL EXPLORATIONS

Small Builds.
Unfinished Ideas.
Useful Experiments.

Experiments where I test a product interaction, architecture, model, workflow, or technical assumption. No polished theater—each entry records what worked, what failed, and what to test next.

TOTAL LOGGED:04 ARCHIVED

RUNTIME ENVIRONMENT:CPU / FASTAPI / ONNX

ACCURACY STANDARD:VERIFIED EVIDENCE

LOGGED EXPERIMENTS (4)PROTOCOL: HONEST FAILURE LABELS

LAB-01/Machine Learning / CPU Inference

Completed Prototype

In-Process ONNX Query Routing & Roman Urdu Detection

HYPOTHESIS UNDER TEST:

"A compact MiniLM or logistic regression classifier exported to ONNX Runtime can reliably predict intent, route (RAG vs. deterministic), and detect Roman Urdu on CPU in under 5ms, eliminating LLM latency and API spend."

WHAT WAS BUILT

Trained a 4-class multi-output classifier in scikit-learn / Hugging Face, exported to ONNX format, and mounted in-process inside FastAPI at application startup.

WHAT WORKED

Sub-4ms classification latency on single CPU core; 100% precision on high-confidence greetings, contact lookups, and clear Roman Urdu queries.

WHAT FAILED

Boundary queries with ambiguous phrasing (e.g. asking about both design and pricing) exhibited lower confidence, requiring safe deterministic fallback routing.

NEXT HYPOTHESIS

Benchmark INT8 quantized transformer models to improve semantic nuance while staying strictly within the 5ms latency window.

Source Code Interactive Demo

DATE: 2025–2026

LAB-02/LLMOps & Evaluation

Completed Prototype

Automated RAG Retrieval Regression Gate in CI

HYPOTHESIS UNDER TEST:

"Automating Hit@5 and MRR evaluation across gold-standard questions in GitHub Actions can block PRs before inaccurate chunking degrades production answers."

WHAT WAS BUILT

A deterministic normalization and Qdrant ingestion script paired with a Python test harness that executes 50 canonical test queries against staged vectors.

WHAT WORKED

Successfully blocked an experimental chunk size modification that caused a 12% regression in retrieval recall for technical case studies.

WHAT FAILED

Synthetic test query generation generated overly simplistic questions that did not mirror the linguistic diversity of real recruiters.

NEXT HYPOTHESIS

Maintain a locked, human-authored evaluation dataset with strict paraphrase leakage prevention across splits.

Source Code

DATE: 2025

LAB-03/Real-Time Voice AI

Active Exploration

Conversational Companion with Server-Enforced Quota Safeguards

HYPOTHESIS UNDER TEST:

"A persistent floating widget with real-time ElevenLabs conversational voice can provide high engagement while server-enforced 120-second session limits and monthly quota degradation prevent unexpected cloud bills."

WHAT WAS BUILT

React companion widget with separate Text and Voice tabs, backed by a FastAPI OpenAI-compatible proxy adapter that connects to the LangGraph RAG pipeline.

WHAT WORKED

Conversational audio streaming with natural turn-taking; clear countdown timer informing the user of remaining seconds; zero client-exposed API keys.

WHAT FAILED

Mobile browser autoplay restrictions require strict user-gesture initiation before audio context initialization.

NEXT HYPOTHESIS

Implement adaptive sample-rate negotiation to improve conversational stability over high-latency cellular connections.

Interactive Demo

DATE: 2026

LAB-04/Document AI & Vision

Archived Experiment

Document Table Extraction for Enterprise Tax Filings

HYPOTHESIS UNDER TEST:

"Pairing OCR bounding boxes with schema-enforced Pydantic extraction can digitize low-quality scanned IRS tax forms with zero manual data entry."

WHAT WAS BUILT

Multi-stage pipeline utilizing vision models for table identification followed by structured JSON extraction using function calling.

WHAT WORKED

Achieved high accuracy on clean digital PDFs and standardized 1040/W-2 schedules.

WHAT FAILED

Low-resolution smartphone scans with non-uniform lighting and crumpled folds caused table boundary hallucination and digit misreads.

NEXT HYPOTHESIS

Pre-processing pipeline utilizing perspective correction and contrast-limited adaptive histogram equalization (CLAHE).

DATE: 2024

Loading content...

APPLIED R&D & TECHNICAL EXPLORATIONS

Small Builds.
Unfinished Ideas.
Useful Experiments.

Experiments where I test a product interaction, architecture, model, workflow, or technical assumption. No polished theater—each entry records what worked, what failed, and what to test next.

TOTAL LOGGED:04 ARCHIVED

RUNTIME ENVIRONMENT:CPU / FASTAPI / ONNX

ACCURACY STANDARD:VERIFIED EVIDENCE

LOGGED EXPERIMENTS (4)PROTOCOL: HONEST FAILURE LABELS

LAB-01/Machine Learning / CPU Inference

Completed Prototype

In-Process ONNX Query Routing & Roman Urdu Detection

HYPOTHESIS UNDER TEST:

WHAT WAS BUILT

Trained a 4-class multi-output classifier in scikit-learn / Hugging Face, exported to ONNX format, and mounted in-process inside FastAPI at application startup.

WHAT WORKED

Sub-4ms classification latency on single CPU core; 100% precision on high-confidence greetings, contact lookups, and clear Roman Urdu queries.

WHAT FAILED

Boundary queries with ambiguous phrasing (e.g. asking about both design and pricing) exhibited lower confidence, requiring safe deterministic fallback routing.

NEXT HYPOTHESIS

Benchmark INT8 quantized transformer models to improve semantic nuance while staying strictly within the 5ms latency window.

Source Code Interactive Demo

DATE: 2025–2026

LAB-02/LLMOps & Evaluation

Completed Prototype

Automated RAG Retrieval Regression Gate in CI

HYPOTHESIS UNDER TEST:

"Automating Hit@5 and MRR evaluation across gold-standard questions in GitHub Actions can block PRs before inaccurate chunking degrades production answers."

WHAT WAS BUILT

A deterministic normalization and Qdrant ingestion script paired with a Python test harness that executes 50 canonical test queries against staged vectors.

WHAT WORKED

Successfully blocked an experimental chunk size modification that caused a 12% regression in retrieval recall for technical case studies.

WHAT FAILED

Synthetic test query generation generated overly simplistic questions that did not mirror the linguistic diversity of real recruiters.

NEXT HYPOTHESIS

Maintain a locked, human-authored evaluation dataset with strict paraphrase leakage prevention across splits.

Source Code

DATE: 2025

LAB-03/Real-Time Voice AI

Active Exploration

Conversational Companion with Server-Enforced Quota Safeguards

HYPOTHESIS UNDER TEST:

WHAT WAS BUILT

React companion widget with separate Text and Voice tabs, backed by a FastAPI OpenAI-compatible proxy adapter that connects to the LangGraph RAG pipeline.

WHAT WORKED

Conversational audio streaming with natural turn-taking; clear countdown timer informing the user of remaining seconds; zero client-exposed API keys.

WHAT FAILED

Mobile browser autoplay restrictions require strict user-gesture initiation before audio context initialization.

NEXT HYPOTHESIS

Implement adaptive sample-rate negotiation to improve conversational stability over high-latency cellular connections.

Interactive Demo

DATE: 2026

LAB-04/Document AI & Vision

Archived Experiment

Document Table Extraction for Enterprise Tax Filings

HYPOTHESIS UNDER TEST:

"Pairing OCR bounding boxes with schema-enforced Pydantic extraction can digitize low-quality scanned IRS tax forms with zero manual data entry."

WHAT WAS BUILT

Multi-stage pipeline utilizing vision models for table identification followed by structured JSON extraction using function calling.

WHAT WORKED

Achieved high accuracy on clean digital PDFs and standardized 1040/W-2 schedules.

WHAT FAILED

Low-resolution smartphone scans with non-uniform lighting and crumpled folds caused table boundary hallucination and digit misreads.

NEXT HYPOTHESIS

Pre-processing pipeline utilizing perspective correction and contrast-limited adaptive histogram equalization (CLAHE).

DATE: 2024

Small Builds.Unfinished Ideas.Useful Experiments.

In-Process ONNX Query Routing & Roman Urdu Detection

Automated RAG Retrieval Regression Gate in CI

Conversational Companion with Server-Enforced Quota Safeguards

Document Table Extraction for Enterprise Tax Filings

Small Builds.Unfinished Ideas.Useful Experiments.

In-Process ONNX Query Routing & Roman Urdu Detection

Automated RAG Retrieval Regression Gate in CI

Conversational Companion with Server-Enforced Quota Safeguards

Document Table Extraction for Enterprise Tax Filings

Small Builds.
Unfinished Ideas.
Useful Experiments.

Small Builds.
Unfinished Ideas.
Useful Experiments.