Selected AI Work

MSc research · Multi-agent · Voice

HireAgent: Multilingual Multi-Agent Job Platform

MSc Data Science final research · Coventry University UK (NIBM) · 2025/26

A conversational AI platform for Sri Lanka's informal employers. They speak or text a job in any local language, and a multi-agent pipeline turns it into a moderated, published listing. No forms, and no English required.

Role
Researcher · AI Engineer
Year
2025 – 26
Focus
MSc research
Stack
8 technologies
  • CrewAI
  • OpenAI
  • MCP
  • FastAPI
  • Twilio Voice
  • pgvector
  • Redis
  • Railway
AgentsIntent gatewayAPIWorkerMCP serverPrototypeViva slides

  • 0.83

    Extraction F1 (P 0.87 / R 0.80)

  • 433ms

    Real-time search latency

  • 1.00

    Scam-detection recall

  • 92.5%

    Intent routing accuracy

  • 95.8%

    Functional tests passed (23/24)

  • 0

    False positives on clean inputs

The problem

About 70% of employment in Sri Lanka is informal. Employers hire by word of mouth, signage, and WhatsApp, while job portals are desktop-first, English-only, and full of forms. Mobile internet reaches 80% of people, but computer literacy is only 36.4% and desktop ownership 19.5%. On top of that, unmoderated job posts expose job seekers to fraud and exploitation.

Overview

I designed, built, and evaluated a multi-agent platform to remove those barriers. It accepts voice calls, SMS, WhatsApp, and web chat. It detects language and intent, asks clarifying questions, and pulls out a structured job posting. A six-stage CrewAI pipeline then validates, enriches, classifies, moderates, and publishes the posting.

The research followed Design Science Research (Hevner et al., 2004) with iterative prototyping: 8 sprints over 14 weeks. I tested it on scenario corpora covering complete English, fragmented Singlish, Tamil-English mixes, Romanized Sinhala, and deliberate scam and discriminatory inputs.

HireAgent began as a single-repository prototype. It grew into a distributed system of four core repositories (published as BrokeMe) deployed on Railway as 7 services.

Architecture

End-to-end architecture

Real-time and asynchronous work are split. Searches return in under 500ms, while AI listing creation (about 7.6s) runs in the background so the user never waits.

  1. 01

    Omnichannel intake

    Voice · SMS · WhatsApp · Web

    Twilio ConversationRelay for live calls, plus SMS, WhatsApp, and web chat.

  2. 02

    Understanding layer

    brokeme-intent

    Language detection, intent detection, and conversational extraction into a structured schema, with session state in Redis.

  3. 03

    Bifurcated routing

    433ms vs async

    Customer searches take the real-time path. New listings go onto a Redis queue.

  4. 04

    Worker

    brokeme-worker

    Three daemon threads consume the queue and hand jobs to the agent pipeline.

  5. 05

    Multi-agent pipeline

    brokeme-agents

    Six CrewAI stages with defensive JSON parsing, so hallucinated output can't corrupt later stages.

  6. 06

    Publish & notify

    brokeme-api

    FastAPI and PostgreSQL with pgvector store the listing, publish it, and notify the user.

How it works

Six-stage agent pipeline

Each stage has one job and a tuned temperature: low for checks, higher for creative writing.

  1. 01

    Validation

    Stage 1 · temp 0.1

    Checks the posting is complete across all 8 schema fields.

  2. 02

    Enrichment

    Stage 2 · temp 0.3

    Adds tags, synonyms, and SEO keywords.

  3. 03

    Classification

    Stage 3 · temp 0.2

    Assigns one of 45 categories.

  4. 04

    Moderation

    Stage 4 · temp 0.1

    Screens for scams and discriminatory language before anything is published.

  5. 05

    Content generation

    Stage 5 · temp 0.4

    Writes professional, ready-to-publish ad copy.

  6. 06

    Image prompt

    Stage 6 · temp 0.4

    Generates a DALL·E prompt for the listing visual.

Components

  • brokeme-intent

    Channel gateway: voice, SMS, WhatsApp, and web intake, language and intent detection, structured extraction.

  • brokeme-api

    FastAPI with clean architecture, PostgreSQL with pgvector, OTP auth, and service and listing management.

  • brokeme-worker

    Consumes the Redis queue and runs AI processing off the request path.

  • brokeme-agents

    The CrewAI six-stage pipeline: validation, enrichment, classification, moderation, content, and image prompt.

  • provider-search-mcp

    An MCP server that gives agents safe, read-only tools: search providers, get details, check slots.

Results

Extraction F1 by language

Higher is better. Code-mixed and romanized input is the hardest.

  • Standard English0.93
  • Partial English0.83
  • Tamil-English0.75
  • Code-mix Singlish0.74
  • Romanized Sinhala0.66

Speech recognition word error rate

Lower is better. Controlled 16kHz audio, 19.1% overall.

  • Clear English7.2%
  • Sinhala21.4%
  • Singlish mix28.6%

Research contributions

  • 01Evidence that a multi-agent conversational design lowers cognitive barriers for low-literacy informal employers
  • 02Empirical speech recognition benchmarks for Sinhala and code-mixed Singlish under controlled conditions
  • 03An architecture pattern that splits real-time search from asynchronous AI processing
  • 04An inclusive AI design framework for multilingual, mobile-first developing economies

Limitations & future work

  • Voice was tested through WebSocket simulation, not a live mobile network
  • Agent tests used mocks, so they validate the architecture rather than the meaning of outputs
  • Recall on discriminatory language is 0.80, because implicit bias is harder to catch
  • Next: a formal usability study with Sri Lankan SME owners, token streaming for sub-200ms voice replies, and speech recognition fine-tuned on local dialects