Skip to content
Ajay KrishnanThe Unprofessional

Legal engineering for law firms

Independent by design

Your firm. Compiled.

I'm Ajay. I build AI around your firm's process, precedents and judgment. You shouldn't have to practise someone else's way to use better software.

Start with one workflow Play with the process

Writing software since 2004. Called to the Ontario bar in 2012.

01 / The working paper

Keep your process. Add the useful part.

The documents. The awkward exceptions. The partner who wants it done a particular way. That's the work—not something to design away. Switch on the assistance and see what stays put.

Authored, illustrative examples. Not live AI. No uploads or client material.

From instructions to advice

Your precedents and research habits. A little less hunting, a little more thinking.

  1. Take instructions

    Define the question, the facts and what the client actually needs.

  2. Work through the sources

    Choose the authorities and precedents. Check what applies.

  3. Form the view

    Weigh the argument, resolve the gaps and settle the draft.

  4. Give the advice

    The lawyer reviews, approves and communicates the advice.

An illustrative source

The draft says notice was sent on 14 March. The record contains no delivery receipt.

The lawyer verifies the sources, settles the analysis and approves the advice.

Output: your memo, with unresolved questions still visible.

Open the right matter, properly

Keep the conversations and checks that make a good file. Cut down the re-keying.

  1. Hear the problem

    Understand the request, the parties and the prospective client's expectations.

  2. Check the fit

    Run the firm's conflicts and acceptance process.

  3. Set the scope

    Agree the work, responsibilities and terms of engagement.

  4. Open the file

    Confirm acceptance and assign the matter to the right people.

An illustrative source

Two related companies are named in the email. Only one is listed on the instruction sheet.

Acceptance, conflicts and engagement remain your firm's decisions.

Output: a checked opening record—not another inbox to babysit.

Find the hard questions before the hearing

Your record. Your argument. A rehearsal that sends you back to the right passage.

  1. Set the record

    Choose the materials and identify the issues that matter.

  2. Build the argument

    Develop the theory, concessions and route through the evidence.

  3. Rehearse under pressure

    Test the argument aloud and decide where it needs work.

  4. Settle the preparation

    Verify the references and decide what to take into the hearing.

An illustrative source

At paragraph 37, the witness says the meeting was on Tuesday. The exhibit is dated Wednesday.

The advocate chooses the strategy and checks every authority and reference.

Output: your hearing notes, with the difficult questions kept in view.

I can’t believe it’s not a lawyer.

It doesn’t need to be. Your firm already has those.

Your judgment is the point. The opportunity is in the work around it: finding the passage, comparing the clauses, preparing the draft, keeping track of what nobody has answered yet.

02 / Evidence, not a pitch deck

I build the software. And the parts a demo skips.

Source boundaries. Review points. What happens when the answer is wrong. These are my own products and engineering work—not a list of law-firm clients.

Commercial-contract research

Blackline

Independent product

Open Blackline

A discussion that stays attached to the contract. Questions refer back to passages; a person's own notes are preserved; marked-up PDFs can leave the application.

The useful pattern: keep the reasoning beside the material and make the output portable, rather than burying everything in a chat history.

Research assistance, not legal advice or approval of a contract.

  • Passage references
  • Preserved notes
  • Annotated PDFs

Oral-argument rehearsal

Greenroom

Invite-only beta

Explore Greenroom

Rehearse aloud with an AI jurist using your brief and record. The exchange retains the questions and the passages behind them, so the follow-up is more useful than a vague impression.

Context access is limited to the relevant matter and conversation. A question can be traced back to the material that prompted it.

Citations locate source passages; they don't independently verify an authority.

  • Live voice
  • Record-grounded questions
  • Retained references

Evidence-linked interviews

Slate

Alpha

Explore Slate

Voice interviews evaluated against the employer's own rubric. Scores refer to that candidate's answers—not another interview, an interviewer's words or a later version of the transcript.

Rubrics are snapshotted for each attempt. Stale transcripts and out-of-scope citations are rejected before an evaluation completes.

Hiring software, not a legal deployment. The evidence controls are transferable; hiring decisions stay human.

  • Rubric snapshots
  • Bounded citations
  • Human decisions

TravelAgent.ai · AI Growth Labs

Auto-Trips

Pre-launch beta

Primary owner of the itinerary generator. First make a creative outline; then resolve its stops against real places using Postgres/PostGIS and BM25.

Candidate-validated place IDs removed a class of invented stops. An evaluation harness and per-generation telemetry expose what still fails.

Travel software in pre-launch beta. Grounding reduces specific errors; it doesn't make a model infallible.

  • Place grounding
  • Evaluation
  • Telemetry

03 / Buy, adapt, or build

You might not need me.

If an existing product fits, use it. A bespoke system earns its keep when the part that makes your firm good is the part a standard product asks you to change.

Buy what already works.

Harvey, Legora and other established tools may be the right choice. Start with the job, evaluate the fit and avoid rebuilding a solved problem.

Build what makes you, you.

Your own precedents, review habits, permissions and handoffs. Custom work is worth considering when those details determine whether people will actually use the system.

The first job is choosing the right approach—not justifying a build.

04 / The person behind the pieces

The Unprofessional

Engineer. Lawyer. Executive. Three professions AI was supposedly going to make redundant.

I'm Ajay Krishnan. I've written software, practised law and run companies. I started building in 2004 and was called to the Ontario bar in 2012. Working across those worlds taught me to notice the gap between a convincing demonstration and something people can depend on.

Today, this is an engineering and consultancy offer—not legal representation.

Engineer
Systems that work beyond the demo.
Lawyer
The exceptions are often the point.
Executive
Someone has to live with the result.
Three disciplines / one point of view
Ajay Krishnan
Ajay Krishnan, in 1,024 pieces. Engineering, law and operating experience—more useful together.

Go on. Take me apart.

The long way here2004–2026

03 / The path here

Not a straight line. A useful one.

Software first. Then law. Then companies where both mattered. The thread is building things and taking responsibility for how they work.

  1. 2025 — now

    AI Growth Labs

    Founding Engineer

    Primary owner of Auto-Trips in TravelAgent's pre-launch beta. Also built Flight Exploration: intent parsing, bounded retries and month-by-month fan-out, with handoff to exact-date search. Flight Exploration remains feature-flagged, not yet in production.

  2. 2024 — 25

    City Lifestyle

    CTO & General Counsel

    Wrote an AI sales qualifier in PHP/Symfony that took over first-contact qualification. The 12-person sales team moved from eight front-line and four back-line roles to all twelve working the back line.

  3. 2022 — 23

    Fishtail

    VP, Trade Technology

    Designed and built the cross-border trade-collateral platform: Rails, PostgreSQL and Hotwire on GKE, with real-time document processing. Ran early pilots with smaller exporters and traders.

  4. 2021 — 24

    Caravel Law

    Technology Counsel · fractional

    Fractional, concurrent with operating roles. AI governance, privacy under PIPEDA and GDPR, and SaaS and IP licensing for companies from seed stage to enterprise.

  5. 2018 — 21

    Lobo Genetics

    Co-Founder & CTO

    Wrote the entire backend in Rails and PostgreSQL. Took the buccal-swab test kit through Health Canada approval and set up the certified processing lab. Acquired by Entheon Biomedical.

  6. 2016 — 18

    Restaurant Brands International

    Counsel, Operations & Technology

    Privacy and PCI counsel for the brands' first mobile order-and-pay apps. Program owner for Tim Hortons Canada's Latte Line and digital menu-board rollouts, from procurement through installation.

  7. 2010 — 16

    Law articles & private practice

    Davies Ward Phillips & Vineberg, then my own practice

    Articles from 2010 to 2012. Called to the Ontario bar in 2012, then ran an employment and technology-law practice through 2016.

  8. 2006 — 08

    Infusion · Barclays Capital

    Tech Lead

    Built Regulation W compliance and structured-credit pricing systems in .NET and Java. Worked with the quants to move their Excel toolkits into grid services. Separately opened Infusion's Dubai office.

  9. 2004 — 08

    Savvica

    Co-Founder

    Built an e-learning platform, first in J2EE and later in Rails. Acquired by Educomp.

Three ways of looking at a problem

  • BASc, Engineering ScienceUniversity of Toronto
  • JDUniversity of British Columbia
  • MBADeGroote School of Business, McMaster University

05 / Start small

One awkward workflow. Let’s start there.

Tell me where the work gets stuck. We'll look at whether existing software, a small adaptation or a custom build makes sense.

  1. Show me the work

    Walk through one recurring workflow and the people who actually do it.

  2. Choose the right fix

    Map the documents, decisions and constraints. Buy, adapt or build.

  3. Prove a small piece

    Agree a bounded pilot and test it against examples your firm can judge.

  4. Put it to work

    Agree integration, handover and ongoing maintenance before expanding.

Write to me Find a time

Please don't send client files or confidential information in a first message.