Tharaka at his desk in a Colombo office, turning to the camera with a smile.

AI agents people actually use.

I’m Tharaka Hettihamu, a product leader in Colombo building voice and chat agents for enterprises, with humans still in charge.

See the workTake over a live call

Take over a live call

A sketch of how my voice agents behave. The agent checks the caller, blocks the card and masks the number in the transcript. The risky step, a new card, waits for a person.

You are the person in charge. Approve the replacement when the agent asks, or press “Take over the call” and answer the caller yourself.

Connecting 00:00 Card services · inbound +94 77 ••• ••21

Transcript

  1. Hi, I lost my card this morning. Can you block it?

  2. I can. I’ve confirmed it’s you from your number. Could you read me the card number?

  3. Sure, it’s •••• •••• •••• 4821.

  4. Done. The card ending in 4821 is blocked. I’ve asked a colleague to approve a replacement.

Six AI products I’ve led

  1. Contact-centre platformLive
  2. AI SDR agentLive
  3. Orchestration and governanceIn development
  4. Support agent with evalsLive
  5. Digital-human kioskPilot
  6. Crop intelligenceProof of concept
Biro drawing: hundreds of phone calls stream into a voice agent, pass a guardrail gate, and reach a supervisor with a hand on the switch.

AI contact-centre platform

Live platform · Telco, banking

Telecom and financial-services clients need voice agents, agent assist, analytics and AI quality checks at 300 concurrent calls, under 1.5 seconds of latency.

  • Owned features and the roadmap; wrote a telco technical proposal and project plan.
  • Conceptualised a model that mixes nine Indic languages to handle complex customer flows for a major Indian bank.
  • Worked on agentic customer experience and AI voice agents for telephony systems.
Biro drawing: a magnifying glass over office towers, a research notebook, then a letter folding into a paper plane caught by a buyer.

AI SDR agent for B2B sales

Live · B2B sales

The old sales-development flow was rigid, asked too many questions and didn’t convert.

  • Product manager and author of the functional requirements and screen specs.
  • Redesigned the workflow into three steps: prospect, research, outreach.
  • Shaped it into a modular multi-agent product that takes a lead from prospecting through research to personalised outreach.
Biro drawing: scattered papers carried by small digital workers through a lock, a checklist and a human approver, ending as neat folders.

Enterprise AI orchestration and governance

In development · Enterprise operations

Enterprises have scattered data and processes, and AI use nobody governs.

  • Shaping the roadmap by aligning design, technical and business stakeholders.
  • Driving discovery and building the demos enterprises need, while managing product operations.
  • Integrating conversational AI into digital workers.
Biro drawing: a chat on a phone and a web widget, a balance weighing an answer against books, a five-row checklist, and a person ready to take over.

Customer-support agent with LLM evals

Live · Customer support

FAQ answers on WhatsApp and a web widget, grounded in scoped knowledge, with hand-off to people.

  • Owned requirements and the roadmap, including a test playground.
  • Defined the Langfuse eval dimensions: FAQ relevance, response quality, edge cases, questions outside the knowledge base, and cost.
  • Used an LLM as judge to evaluate every response.
Biro drawing: a visitor in a winter coat talks to a digital human on a tablet kiosk in a Nordic town-hall lobby, snow outside.

Digital-human service kiosk

Pilot · Public sector

A tablet avatar for a Nordic municipality, with face, gaze and gesture detection, and Norwegian voice and chat.

  • Product and project coordination.
  • Built the Norwegian knowledge base, with separate knowledge for visitors and staff.
  • Wrote UI specs and architecture docs; internal testing reached 18 of a 30-person target.
Biro drawing: Sri Lankan tea hills scanned by a drone and a satellite, with a magnified leaf showing disease spots and a water droplet.

Crop-intelligence AI

Proof of concept · Agriculture

Plantation groups need to predict crop health, disease, water stress and yield from satellite and drone imagery.

  • Shaped the solution, with predictive disease management to catch outbreaks early and stop them spreading.
  • Added yield and financial forecasting alongside it.
  • Mapped 26 possible KPIs, working towards disease detection, water-stress monitoring, yield prediction and financial forecasting.

From jet engines to AI agents

  1. TharakaOttawa · 2010

    I trained as an aircraft engineer in Canada. I genuinely wasn’t a very good one.

  2. TharakaThe turn

    I didn’t want to build things that sat in a room.

  3. TharakaColombo · now

    So I build products people use. If one fails, it’s on me.

    How I build a product, drawn as the engine I trained on. Move across it to open the throttle; keep scrolling for what comes out.

    If the product fails, it’s on me.

    Tharaka HettihamuTharaka HettihamuHow I work

    I build agentic and conversational AI for enterprises from Colombo: voice and chat agents, AI SDRs, and the governance that keeps them safe to use.

    I care about the unglamorous parts that decide whether AI works: clear ownership, real-world testing, and value you can actually measure.

    Agents should act, not just chat. I design multi-step workflows with clear guardrails, cited reasoning and human review where it matters, and I measure them with real evals rather than demos.

    Voice is where it gets hard: latency, turn-taking and trust decide whether people keep talking to an agent.

    Before agents I owned Drawify, a visual-storytelling product that made Slush 100, and worked on Spectrify AI, which brings near-infrared quality testing to Sri Lanka’s tea industry. The constant is taking a fuzzy idea to shipped scope, with sharp PRDs and a team that agrees on what matters first.

    Episode cover: Tharaka in conversation with the host of The Creative Odyssey Podcast. Listen: He Builds Things That People Actually Use The Creative Odyssey Podcast · May 2026 · 23 min

    Flight log

    1. 2026Speaker, AI Community Sri LankaHow AI is changing product work, at Microsoft Sri Lanka
    2. 2026Guest, The Creative Odyssey PodcastHe Builds Things That People Actually UseWatch
    3. 2026Judge, enterprise AI hackathonColombo
    4. 2024Spectrify AINear-infrared quality testing for Ceylon tea; one of ten startups in the GoviLab accelerator
    5. 2023Drawify picked for Slush 100Product owner of a visual-storytelling platform built in Colombo
    6. 2023Quoted in Rest of WorldHiring developers through Sri Lanka’s 2022 crisisRead
    7. 2023Product Hunt Launch GuideA practical guide to launch day, on MediumRead
    8. 2010Carleton University, OttawaAircraft engineering on an entrance scholarship; later president of the Sri Lankan Student Alliance

    Compare notes.

    Building something with AI agents, or trying to get one past the demo stage? I’m always happy to compare notes.

    Message me on LinkedIn

    Colombo, Sri Lanka · GMT+5:30 local time

    Drawing of Colombo from Galle Face Green: the Lotus Tower, high-rises, the Galle Face Hotel, palms, kite flyers and the ocean.