Build the system.
Own the decisions.

These results come from recorded engineering checks. Each comparison keeps its method and scope attached.

1. 01Compared

   Isolate the loop cost before replacing the browser library

   I compared execution paths using equivalent browser-command protocols. The library comparison alone did not explain the delays. The investigation instead separated network round trips, repeated observations, model turns, and readiness waits.

   Browser Use · CDP · action-group experiments

   Scope of the result

   A faster executor can still sit inside a slower agent loop. End-to-end runs with different decisions do not establish a causal speedup.

2. 02Development deployed

   Use native framework components, then test the behavior

   I compared the existing agent design with native framework composition and revised the integration around real agent, model, tool, and typed-output components. Shared runtime helpers stayed focused on execution checks.

   Agno · Python · typed outputs · development smoke tests

   Scope of the result

   Local tests and development deployments are recorded. A change in framework or model cost is not, on its own, a demonstrated behavior or latency improvement.

3. 03Historical experiment

   Explore GPU-backed speech serving

   I explored GPU-backed speech generation using Modal, vLLM, and Higgs Audio. A short-lived shared prototype was removed; a separate historical speech-serving implementation followed. The work covers serving integration, rather than authorship of the underlying model.

   Modal · vLLM · PyTorch · HiggsAudio

   Scope of the result

   Historical implementation evidence exists. Current deployment, throughput and voice-latency gains were not verified.

4. 04Prototype

   Prototype faster projection writes while preserving recovery

   I evaluated batched Redis writes against a Lua-based atomic path, with recovery checks alongside latency measurements. The prototype remained outside the merged delivery path.

   Redis · Lua · PostgreSQL · recovery tests

   Scope of the result

   The experiment used a local database and remote Redis with synthetic notification delay. It is not a customer-performance result.

5. 05Development acceptance

   Build a device fleet around actual demand

   I worked on per-device allocation and scale-to-zero instead of keeping a large shared host warm. The recorded development acceptance tested ten simultaneous leases, repeated acquisition, the capacity ceiling, and cleanup back to zero.

   GCP · Android execution · idempotent leases · crash recovery

   Scope of the result

   The session reports development acceptance. Extended workload stability and production promotion remained separate gates. Cost comparisons were models, not settled invoice savings.

6. 06Investigated

   Make benchmark integrity a prerequisite for a score

   I investigated navigation benchmarks where dataset preconditions shifted step evidence and invalid judge responses affected eligibility. The work separated harness failures from agent failures and preserved historical scorecards while defining a corrected verification cohort.

   Navigation evaluation · evidence provenance · screenshot judging

   Scope of the result

   Invalid or incomparable runs are not treated as improvements in agent accuracy.

7. 07Lab comparison

   Test a model change before changing the product

   I tested an alternative model with explicit prompt caching against the current exploration policy. In one three-run comparison, the candidate’s median journey took 44.8 seconds versus 41.3 seconds for the baseline; all six runs completed correctly. The small sample did not justify a model switch.

   Python · OpenRouter · prompt caching · browser-agent evaluation

   Scope of the result

   October 2026 lab runs on one login-form journey. Provider timing and agent decisions vary; this is a recorded comparison, not a general ranking of models. Cost observations came from separate instrumented runs.

8. 08Exploratory lab comparison

   Separate infrastructure delay from agent behavior

   I built an experiment that changed database placement and simulated orchestration delay around the same agent version. One run took 124.8 seconds with the delayed setup; one took 76.7 seconds with a regional database and no added orchestration delay. Both passed the lab checks.

   PostgreSQL · Toxiproxy · cloud browsers · Python

   Scope of the result

   One run per setup. Database, orchestration delay and browser waits differed, so this observation does not isolate a database-only gain or establish a production speedup.