Software efficiency for AI-heavy products
I help AI-heavy SaaS and software teams reduce infrastructure, model/API, database, and workflow costs without compromising product quality. Every change is tested against the outcomes your users need.
30 minutes with Fady. Free initial assessment of 1–3 areas worth investigating.
No codebase access, API keys, or credentials needed.
How I can help
For AI-heavy software teams facing rising model bills, slow queries, or repeated processing. Infrastructure and workflow costs matter alongside model choice.
We start with your costs, architecture, and the issues you’ve noticed. No codebase access, API keys, or credentials needed.
You get A prioritized shortlist of around 1–3 high-value areas worth investigating, why they may be inefficient, and what to test first. These are hypotheses, not guaranteed savings. Further work starts only once we agree on scope and cost.
Test one change against the same workload, with quality and correctness checks agreed upfront.
You get A before-and-after comparison, documented tradeoffs, and a keep-or-reject recommendation.
Put a supported improvement into your codebase and make it clear how to verify and maintain it.
You get A focused code change, relevant checks, and rollout and rollback guidance.
The portfolio / Independent open-source work
Database queries, document processing, and AI workflows. See what changed, what the tests show, and where the limits are.
Database performance
Quota checks scanned chat records. A targeted database index made the tested busy-user query faster, with slower writes and extra storage as the tradeoff.
Busy-user quota query · 1M rows · 8 MiB cache
Median of three per-trial medians
Counting a user’s recent messages required scanning chat records. History queries needed those records filtered by conversation and returned in order. An index optimized only for the count could make history reads slower by introducing a separate sort.
Use a composite index arranged around the user and conversation filters, followed by ordering and date fields. Quota counts can read the smaller index entries, while matching history queries can traverse them in order. The date still needs checking within the user’s entries.
The tested large workloads with quota checks benefited. The index also adds about 9% storage and slows inserts. At a million rows, every tested sequence without the quota check was slower, making workload fit essential.
Evidence limits. Measured locally on synthetic data. Whole-app speed and bill savings not measured.
Document intelligence
Five separate requests repeatedly processed the same document. Keeping the title first and grouping the remaining fields reduced that repetition in the synthetic test.
Title, tags, correspondent, document type, and date were requested separately. That repeatedly sent the same document through the model. Later fields also depend on the generated title, so simply merging all five requests would lose useful context.
Keep the title request, then ask for the other four fields together. Require a complete structured response and validate its fields. If grouping fails, return to the original individual requests. Custom prompts and other excluded configurations retain their existing behavior.
The fixed synthetic example returned matching metadata with five requests reduced to two. A failed grouped request followed by successful fallback takes six, so the real benefit depends on how often grouping succeeds.
Evidence limits. Tested locally with simulated responses. Live accuracy and billed savings not measured.
An incompatible voice could derail a podcast after generation had started. A small offline check now catches known mismatches before that work begins.
A known failure becomes a useful early answer.
Podcast creation is a sequence: generate an outline, write the transcript, then create the audio. A voice could be present in the configuration but incompatible with its speech model. Discovering that only at the audio stage wastes the earlier work.
Validate the selected model and voice together before generation starts. Check each speaker’s override as well as the default. A local catalogue handles the two known Gemini models without adding a network request; unfamiliar models keep their existing route.
A known mismatch produces an actionable error before the outline, transcript, or audio is attempted. The user can correct the voice setting first.
Evidence limits. Verified with local tests. Live savings and speech quality not measured.
Opportunities under evaluation
These estimates guide what to test. They are not measured savings: creative quality and successful caller outcomes still need live evaluation.
Creative generation
Reuse eligible creative assets, render simple motion locally, and track the cost of every attempt. A more deliberate workflow reduces the need to generate every variation from scratch.
Eight-format campaign + two motion outputs
Generating each platform format independently repeats image and copy work. Simple camera-like motion also goes through a generative-video service, even when it only needs a predictable transformation of an existing image.
Separate reusable visual masters from platform exports. Use local rendering only for motion that does not need new action, geometry, or viewpoints. Track operations and attempts durably so retries, provider progress, and completion remain distinguishable.
Five reusable masters, one grouped copy request, 60–70% eligible local motion, and an allowance for local compute. Reuse still needs to preserve each format’s creative purpose. The broader implementation is local; the separate public PR fixes provider polling.
Evidence limits. Implementation checked locally. Savings are modeled; creative quality needs live evaluation.
Inspect the polling fix ↗Realtime voice
Make a lower-cost voice model an explicit option, then evaluate it against the whole call: model usage, telephony, retries, and a successful caller outcome.
Depends on workload and comparable caller outcomes
The realtime model is selected when a voice session is configured. Using a less expensive model can lower audio-processing costs, but telephony and session charges remain. Comparing model prices alone exaggerates the potential whole-call saving.
Update the supported model options and make the smaller model an explicit configuration choice. Keep the full model as the default. Calculate both model and non-model costs, and evaluate transfers, interruptions, retries, and successful caller outcomes before deciding to switch.
The planning range uses a fixed speech workload and different non-model cost assumptions at September 1 prices. It depends on the smaller model preserving caller outcomes; the configuration change alone does not establish that.
Evidence limits. Configuration tested locally. Savings are modeled; live call quality not measured.
Inspect the implementation ↗* Contribution statuses are a snapshot from 15 September 2026. All five upstream proposals were open and unmerged.
How I work
The unit that matters is a successful user outcome. Failed attempts, retries, slow responses, and engineering effort belong in the calculation.
Agree on the outcome and the quality that must be protected.
Compare the same workload and keep the evidence inspectable.
Record the benefit, the tradeoffs, and the reasons to keep or reject it.
The person behind the work
I’m Fady, the person behind Finups, an independent software optimization consultancy. I investigate where software does unnecessary work, then test whether a focused change improves its cost or performance.
My open-source work spans database performance, document processing, and AI generation. You can inspect the changes, the tests, and the tradeoffs in each case.
Meet the work on GitHubStart with one workflow
Start with a free 30-minute conversation with Fady. Get an initial assessment of around 1–3 high-value areas worth investigating, why they may be inefficient, and what to test first. No codebase access or API secrets needed. No guaranteed savings or obligation to continue.