Overview
Hamming AI is an active testing and monitoring company for voice and chat agents. Sumanyu Sharma and Marius Buleandra incorporated the company in December 2023, entered Y Combinator's Summer 2024 batch, and initially pursued broad AI evaluation and trust infrastructure. After roughly eight or nine voice-agent teams described the same manual call-testing problem within two weeks, Hamming narrowed to voice QA.[1]
The product now spans simulated pre-launch calls, production monitoring, red-teaming, production replay, audio-native metrics, and CI gates.[2] The strategic asset is a production-to-regression evidence loop across voice runtimes. The risk is category compression: Cekura and Coval compete directly, while voice runtimes can bundle testing. Hamming is not a failure story. YC marks it active, the company is hiring, and its blog published in 2026.[3]
Founding Story
YC identifies Sharma as founder and CEO and says he previously led data at Citizen and worked as a senior staff data scientist at Tesla. The profile attributes large user and revenue outcomes to those roles, but those figures were not audited in this research.[3] The August 2024 Launch HN introduction names Sharma and Buleandra as founders. It says Buleandra ran data infrastructure at Anduril, worked with Sharma at Citizen, and was a founding engineer at Spell before Reddit acquired it.[4]
The founders began with a general reliability thesis. A 2026 retrospective traces Sharma's concern to alleged silent data corruption at Tesla that reduced conversion by 15% and took 72 hours to diagnose, plus Citizen's need to review noisy public-safety audio. These are founder accounts, not independently reconstructed incidents.[1]
For months the company lacked a sharp wedge. A June 2024 Product Hunt newsletter still described a broad benchmark and reliability product for AI systems.[5] Around June or July, eight or nine voice-agent teams independently described manual call testing. The repetition prompted the July voice-QA pivot.
The canonical format requests two exact founder quotations here. The final research preserves founder claims and full-source URLs, but not exact quoted wording. Reconstructing speech would be unsupported, so the two-quote gap is explicit.
The name references Richard Hamming and the Hamming-distance metaphor. Sharma said Hamming's essay “You and Your Research” influenced the team and that he rereads it quarterly.[4] More practically, Tesla's simulation approach influenced the product: virtualize an environment, change the system, and test before failures reach users.[6]
The current founder display is inconsistent. Launch and financing materials identify Sharma and Buleandra, while the current YC interface duplicates Sharma and does not clarify Buleandra's current role. The evidence supports the founding pair, not a conclusion about Buleandra's present operating status.
Timeline
- December 2023: Hamming incorporated with a broad AI evaluation thesis.[1]
- June 2024: Product Hunt documented its pre-pivot benchmark and reliability positioning.[5]
- July 2024: Repeated customer interviews prompted a dedicated voice-agent QA product.
- August 15, 2024: Hamming launched automated voice-agent testing on Hacker News.[4]
- September 2024: TechCrunch selected Hamming among 13 notable companies from the first day of YC Demo Day.[7]
- December 2024: Hamming announced a $3.8 million seed round led by Mischief.[8]
- June 2025: Hamming announced testing and monitoring support for Webex AI Agent customers. Cisco confirms that Webex AI Agent exists but does not independently name Hamming.[9][10]
- 2026: Hamming published its product-history retrospective, remained live, and continued hiring.[11]
What They Built
At launch, Hamming generated personas and scenarios, placed automated agent-to-agent calls, injected noise, silence, and interruption, then scored results with deterministic checks and LLM judges. It could also evaluate production conversations, while production replay, scenario generation, and automatic improvement remained roadmap items.[4]
Voice created a distinct QA surface. A text agent can be checked for content and tool use. A voice agent must also handle latency, barge-in, interruption timing, silence, background noise, accents, monologues, DTMF tones, IVR navigation, and speech-recognition errors. Hamming's current product advertises more than 50 metrics, more than 65 languages, and tests for those audio conditions.[2]
The platform expanded in both directions. Before launch, teams simulate scenarios and adversarial conditions. After launch, they monitor production calls, find recurring failure patterns, replay production failures as regression tests, and gate releases in CI. Integrations with LiveKit, Pipecat, ElevenLabs, Retell, Vapi, and Hopper position Hamming above voice runtimes rather than inside one.
Hamming says its simulator has greater than 95% agreement with production outcomes and supports more than 50,000 concurrent calls. It reports more than four million calls tested and more than 10,000 agents monitored.[2] These are unreplicated company claims. At launch, the founders said customer data was isolated, not sold, and not used for model training. The current site advertises SOC 2 Type II compliance and HIPAA support with a BAA.[4]
The deepest product loop is production-to-regression evidence. A production failure becomes a reusable scenario; the simulator reproduces relevant conditions; evaluators judge the candidate fix; a release gate records whether known failures remain fixed. That history can outlast any one underlying voice runtime.
Market Position
Target Customers
Hamming targets teams shipping voice agents in customer support, healthcare, sales, scheduling, and other operational workflows. The homepage displays Lorikeet, Luma Health, Netomi, Maven AGI, and Ellipsis Health logos, but does not disclose contract scope or whether every logo is a current paying customer.[2]
Market Size
No reliable voice-agent QA market size was found. The $3.8 million seed and competitor activity show investor and vendor attention, not category revenue. Public pricing is also unavailable, preventing a bottoms-up estimate.
Competition
Cekura offers overlapping pre-launch evaluation, production monitoring, scenario generation, parallel calls, interruption testing, replay, custom metrics, and feedback loops.[12] Its LiveKit page claims more than five million tested minutes and 60,000 calls per day, vendor-reported scale that confirms an active category.[13]
Hamming names Coval as a direct competitor and differentiates on voice-specific adversarial tests, regression gates, replay, and monitoring. That comparison comes from Hamming, not a neutral benchmark.[14] Voice runtimes are another structural competitor because they can bundle testing around their own transports, models, and telemetry.
Cross-runtime neutrality is Hamming's counter. A QA layer can compare providers, preserve regressions through migrations, and test host-application behavior. Its weakness is dependence on simulator realism and evaluator reliability. A 2025 research preprint argues that buyers must assess both, meaning the testing platform itself requires validation.[15]
Launch discussion also challenged the category from below. Commenters asked whether ordinary self-service interfaces or simpler intent-based bots were enough and raised labor-displacement concerns. The founders argued that testing should encode criteria from both engineers and domain experts.[4] That debate matters because QA does not decide whether a voice agent should exist. It only measures behavior against declared requirements. Domain policy and human accountability remain outside the simulator.
Business Model
At launch, the founders described pricing as a combination of usage and seats. The current FAQ says startups and SMBs pay by test volume while enterprises receive custom plans, but publishes no prices.[16] Revenue, ARR, retention, gross margin, contract value, burn, and runway remain unknown.
The financing announcement claims automated testing is 20 times faster and 10 times cheaper than human testing, without an independent benchmark.[8] The business must balance usage revenue against simulation and model-evaluation cost while proving that automated results deserve release authority.
Mischief investor Lauren Farleigh framed the round around a widening gap between conversational AI capability and testing, governance, developer, and compliance tooling. Axios independently covered the financing and tied investor interest to production reliability.[18] Those sources confirm financing and the investor thesis, not customer economics.
Traction
Hamming-hosted evidence claims Podium runs more than 5,000 scenarios monthly across eight-plus languages, reduced manual testing 90%, and detected a 15% decline in French appointment recognition before wider rollout. It also reports Scottish-accent recognition rising from 30% to 95% and voice CSAT increasing 12%. None was corroborated on a Podium-controlled source.[17]
Independent signals include TechCrunch's Demo Day selection and Axios coverage of the seed round.[18] Current operation is stronger: YC marks Hamming active, lists two open engineering roles, and the live blog contains 2026 posts.[3]
YC lists the open product-engineering and backend-infrastructure roles at $140,000 to $200,000 with 0.05% to 0.20% equity and possible locations in San Francisco, London, or Austin. Hiring terms are an operating signal, not evidence of revenue or runway.
Post-Mortem
The failure premise is false
Hamming is active, hiring, live, and publishing. There is no shutdown, acquisition, decline, or terminal outcome in the evidence. The canonical format requests a named direct quote supporting the outcome, but the prepared fact corpus preserves no exact founder quote text. No quotation is invented.
The pivot found a real QA surface
The company began broad and spent months without a wedge. Repeated voice-team interviews concentrated the product around a manual testing bottleneck. Hamming then co-developed evaluations with early customers rather than waiting for a complete self-serve product.[19] This sequence supports a product-discovery strength: narrow after repeated demand, then learn inside real workflows.
The evaluator becomes part of the risk
Automated QA scales only if simulated callers resemble real users and automated judges agree with competent humans. Accent, noise, interruption, latency, DTMF, tool use, and high-risk policy behavior create many ways to pass synthetic tests while failing production. The 2025 preprint makes this a structural concern, not an edge case.[15]
Hamming's production monitoring and replay reduce that risk by feeding observed failures back into regression suites. Yet company-reported simulator agreement and customer results do not independently prove calibration across domains. Regulated buyers may require human and expert adjudication before trusting an automated release gate.
The strongest counterargument is that perfect external calibration is unnecessary if the product reliably catches regressions relative to a customer's own baseline. That relative value can be real even when automated scores are not universal truth. The risk returns when procurement, compliance, or release teams interpret vendor metrics as objective safety evidence rather than bounded tests.
Competition can compress the category
Cekura and Coval attack the same workflow, while runtime vendors can bundle basic simulation and monitoring. Hamming must own the cross-runtime evidence history, adversarial corpus, and production-to-regression loop. If testing remains a runtime feature, an independent QA vendor loses pricing power. If enterprises need neutral evidence across providers, specialization becomes more valuable.
Key Lessons
- Repeated customer pain beat a broad thesis. Eight or nine similar conversations in two weeks gave the company a narrower product than months of general evaluation work.
- Audio makes QA multidimensional. Content correctness is insufficient when latency, interruption, accents, noise, DTMF, and tool execution shape the result.
- Production failures compound the test asset. Monitoring and replay turn incidents into persistent regression evidence that can survive runtime changes.
- The judge must be judged. Simulator agreement and automated scores require human calibration, especially when release gates affect regulated workflows.
Sources
- Hamming product-history retrospective
- Hamming
- Y Combinator company profile
- Launch HN
- Product Hunt newsletter
- Deepgram AI Minds interview
- TechCrunch Demo Day selection
- Seed financing announcement
- Hamming Cisco partnership announcement
- Cisco Webex AI Agent documentation
- Hamming blog
- Cekura
- Cekura LiveKit page
- Hamming versus Coval
- Voice-AI testing preprint
- Hamming FAQ
- Podium case study
- Axios financing coverage
- Frontlines founder interview