GitHub projects — Priyam Sameer

01

AI, voice & lending systems

Original interfaces, explicit model paths and deliberate human control.

Lending workspace · three related prototypes

EduLoans

01

Original interaction prototype — sample data; scripted responses.

Student intake

Partner pipeline

Reviewer queue

Inside the MCP prototype

Agent clientRequests a defined tool
MCP contractJSON Schema · stdio
Rules & lookupDeterministic result

Evaluation explains a pre-check. It does not approve a loan.

From a student conversation to an inspectable lending pre-check. One project, with separate interfaces for intake, operations and callable policy.

MCP SDK
JSON Schema
React
Supabase
Deterministic BRE

Human review is a product boundary

Student intake, partner handoffs and reviewer inspection share a case. Uncertain documents and unconfirmed thresholds remain visible before a reviewer acts.

Give AI a tool contract, not lending discretion

The MCP prototype exposes nine typed tools over stdio: profile evaluation, score explanation and institution lookups. The rule engine owns the answer; an agent calls a defined capability.

Keep versions inspectable

The interaction prototype, Supabase matching implementation and MCP server are separate experiments. Their source is grouped here without implying one connected production system.

Documented MCP scoring: S = 0.10U + 0.70A + 0.20C
University, academics and co-applicant components. Deterministic pre-check; the lender retains the credit decision.

The nine MCP tools

evaluate_profile
check_university
check_pincode
list_lenders
score_breakdown
get_scoring_rules
search_universities
list_countries
list_courses

Original interaction prototype — sample data; scripted responses. MCP and matching code are prototypes. Archived March 2026 validation records 145 assertions; not rerun here and includes mocks.

Voice, retrieval & generation

THEYA.Audio

02

THEYA original two-voice RAG Studio interface — Original studio interface · synthetic input; services disabled.

Turn a document into a two-voice listening experience, with explicit choices about retrieval, dialogue and speech synthesis.

FastAPI
Pydantic
Kokoro-82M
Edge Neural TTS
RAG

Separate what is said from how it sounds

Typed sentence records carry display text, speech text, speaker, voice and pauses. Extraction, dialogue generation and speech rendering are separate stages.

Make the model path explicit

Provider adapters in source cover Qwen2.5-72B, Gemini 2.0 Flash, Llama 3.3 via Groq and GPT-4o-mini for dialogue. Speech routes to Edge Neural or Kokoro-82M through Hugging Face/fal-ai. Each path needs provider configuration.

Retrieve before answering; inspect the checks

Overlapping document chunks live in SQLite. BM25, TF-IDF cosine and bigram scores retrieve passages. Grounding, term coverage and dialogue checks are heuristic diagnostics; displayed scores are not measured model accuracy.

Source architecture: document → retrieval / dialogue → speaker turns → TTS → streamed audio
Optional provider integrations exist in source; no inference quality or latency benchmark is claimed.

Original 2-Voice RAG Studio interface with synthetic dialogue input. Backend, model and audio services disabled for this preview.

Telephony & structured extraction

Conversational screening

03

Original screening campaign dashboard using synthetic demonstration records — Original dashboard · sample records; telephony and AI services disabled.

Carry a campaign’s questions into a voice session, then turn the conversation into structured campaign records.

Twilio
ElevenLabs
WebSocket
Gemini 2.5 Flash
Supabase

Preserve context across the call

A Supabase function initiates the Twilio call. A TwiML handler and WebSocket bridge connect it to ElevenLabs conversational AI with campaign context.

Extract after the conversation

A separate function calls Gemini 2.5 Flash through the Lovable gateway with a structured extraction schema. The prototype also maps recommendations into candidate statuses; those assignments have not been validated as hiring decisions.

Documented call path: campaign → Twilio → ElevenLabs session → Gemini extraction → stored record
The UI and integration code are inspectable. No hiring accuracy or completed-call benchmark has been established here.

Original React dashboard with synthetic records. No calls, candidate data or model services connected.

02

Lending & financial decision systems

Callable policy, deterministic rules, operational workflows and explicit financial models.

Fintech comparison & recommendation

Rocket Card Finder

01

Compare cards against spending patterns, fees and eligibility inputs, with the savings calculation visible at the decision point.

React
TypeScript
Vite
Card APIs
Scenario modelling

Ask about spending before ranking

The questionnaire captures category spending; the interface separates total savings from net savings after fees.

Keep comparisons actionable

Bank, network, fees and category filters sit beside card comparison. The source calls BankKaro UAT APIs; the isolated preview disconnects them and shows the original spending-input journey.

Implemented product journey: spending inputs → calculation API → savings response → comparison
Calculation depends on the service. No LLM integration was located; lender eligibility and product benefits require verification.

Original React interface · illustrative spending input; financial APIs disabled. MoneyControl branding is part of the source UI; an organizational relationship is not established.

Exit economics

EquitySim

02

Make liquidation preferences tangible: change an exit value and see what remains for a common shareholder.

JavaScript
Chart.js
Scenario modelling

Show the consequence immediately

Exit value, funding, preference multiple and ownership feed one payout and a doughnut allocation chart.

Implemented simplified model: max(0, exit − funding × preference) × ownership
Preference-first common pool illustration. Excludes preferred conversion, dilution and tax; not a complete cap-table waterfall.

Working browser tool · illustrative inputs; calculations run locally.

03

Local-first developer tools

Structural inference, streaming transformation and visual debugging in the browser.

Structural code generation

JSON Atlas

01

Turn a JSON example into TypeScript interfaces, Python dataclasses or Go structs, without sending the document anywhere.

Recursive inference
Type registry
Three code emitters

Infer once, render three ways

A recursive walk records object shapes in a type registry. Language-specific emitters produce code from the shared structure.

Compiler-style pipeline: JSON → parse → infer structure → type registry → TS / Python / Go
Example-based inference, not schema validation. Array inference begins with the first item; mixed-array unions need review.

Working browser tool · local parsing and code generation; no model/API required.

Visual debugging

Magic Text Matcher

02

Make a regular expression readable through its effects: edit the pattern and watch matching text light up.

ECMAScript RegExp
Capture groups
Escaped highlights

Shorten the feedback loop

Pattern, flags and source text stay on one screen. Match highlighting updates in place instead of asking the reader to inspect raw offsets.

Matching loop: pattern + flags + text → matchAll → escaped highlight overlay
Non-overlapping capture highlights; nested-group rendering is limited. Browser regex rules apply.

Working browser tool · text stays in the tool; no backend or AI inference.

Browser data pipeline

StreamCSV

03

Transform a local CSV into downloadable JSON, with file progress visible as chunks are read.

File.stream()
TextDecoderStream
Blob export

Read incrementally, expose the limits

The browser decodes chunks and updates processed-byte progress. Output rows accumulate in memory before download.

Transformation pipeline: file stream → decode → split rows → JSON chunks → Blob
Simple comma splitting; quoted CSV fields are unsupported. Incremental input does not mean constant-memory output.

Working browser tool · local file processing. Screenshot uses 250 synthetic rows; no large-file benchmark.

04

Creative browser experiences

Procedural generation, reproducible game mechanics and interaction-led discovery.

Procedural exploration

Colour palette generator

01

Original five-swatch colour generator with hex values — Original tool · randomly generated colours.

Generate five colours in one gesture and copy a hex value directly from its swatch.

Random hex generation
Clipboard fallback
Keyboard input

Keep the interaction in the colour

Space generates a palette; a swatch copies its value. Independent random draws provide variety rather than a claimed colour-harmony model.

Per-swatch generation: floor(random × 16,777,215) → hexadecimal → pad to six digits
Procedural colour exploration; no AI or perceptual-harmony algorithm.

Working browser tool · Space to regenerate; select a swatch to copy.

Channel-based discovery

TV Surf

02

Recreate channel surfing with a small curated video library, random starts and a brief tuning transition.

Channel state
YouTube embeds
Keyboard navigation

Let discovery feel like a remote control

Up/down controls wrap through channels. A random selection and start offset change what the next channel reveals.

Original interaction logic: channel ± 1 → wrap → random video + start offset → tune
The isolated preview demonstrates channel controls only. Third-party video playback is disabled; inspect the source for the original embed integration.

Original controls preview · external playback disabled. No video autoplays on this portfolio.

Reproducible daily play

Scrambl

03

A daily word challenge with a shared seed, a ten-second clock and a result designed to be compared.

Mulberry32 PRNG
Fisher–Yates shuffle
Date seed

Make randomness reproducible

The local calendar date seeds the pseudo-random generator. The same date produces the same 18-word sequence.

Daily puzzle construction: YYYYMMDD → seeded PRNG → shuffled 18-word sequence → 10s rounds
Local-calendar seed; not a server-synchronized global midnight. No model calls or account required.

Working browser game · keyboard and on-screen letter controls.

05

Build systems & exploration

Publishing infrastructure, build provenance and the starting points for future experiments.

Publishing & discovery

This portfolio

01

Priyam Sameer original portfolio landing page with painting and agriculture AI fintech headline — Current portfolio interface · original painting.

A Next.js publishing system connecting source-backed case studies, original media, public repository discovery and a CV request flow.

Next.js
React
PostgreSQL
Structured SEO

Keep the evidence attached

The work registry preserves case routes, media and source qualifications. GitHub discovery groups related repositories while retaining each source link.

Treat publishing as a system

Server-rendered pages, canonical metadata and a sitemap make work discoverable. CV requests are recorded before the exact uploaded PDF is delivered.

Publishing flow: authored registry + public repositories → grouped pages → accessible media
Public GitHub data only. Private repositories and backend credentials are excluded from this showcase.

Live portfolio · screenshot captured from the local production build.

Get my CV.

Contact Priyam.

Projects, in practice.